Memory Pulse

memory-pulse

[npm] [license] [MCP]

Site: pulse.strategic-innovations.ai, where the guard runs in the browser. Compare with Mem0, Zep, Letta and CLAUDE.md. Guide: docs/GUIDE.md, the day-to-day protocol with runnable examples.

Give any agent self-improving memory you can trust in production.

Every correction is kept, put in front of your agent when it matters, and enforced. Memory security is built in: a stored note that reads like an instruction never reaches your agent, every row is hash-chained so any edit to the record shows, and a guard blocks the edit that writes a retired value back.

Your agent acknowledged the correction, then wrote the old value back next session. memory-pulse records it once, puts it at the top of every session, and blocks the edit that reintroduces it, quoting the line that retired it. For Claude Code, Codex, Cursor, any MCP client and GitHub Actions. The record is a file in your own repository.

Most memory tools stop at showing the correction to the model. The guard is the part that acts on it.

The four tool definitions come to 3,406 chars, 3.3 KB, about 900 tokens at 3.8 chars per token, measured 2026-09-09. test/client.test.js fails the build past 4,000 chars or 60 tokens either side of 900. Independent measurements put a typical five to ten server MCP setup at 50-67k tokens of tool definitions before your first prompt, roughly a third of a 200k context window.

What it does

A session ends and everything it learned goes with it. memory-pulse keeps a ledger of cause -> effect events in a local file, and gives the agent four tools over it:

Two design decisions do the heavy lifting.

Corrections come first, always. An event recorded with kind: "correction" outranks everything at every brief size and never decays. The failure this prevents is your agent confidently quoting the benchmark number you withdrew three sessions ago.

A weak answer returns nothing. When recall cannot clear its confidence floor it returns an empty result, and the brief says so.

Install

Every route, host by host, with what runs on each: docs/INSTALL.md (also at pulse.strategic-innovations.ai/install).

Six routes. Pick the row for your agent. Node 18 or newer (package.json engines); the test suite runs on 22 and 24.

remember and help from the shell arrive in 0.5.9. npm served 0.5.8 on 2026-09-09, where both exit 1 with unknown command. If npx serves an older cached release, pin this one: npx -y [email protected]; from a checkout, run node server.mjs. Four more things in this README are 0.5.9 behaviour and differ on 0.5.8: guard --verbose names non-JSON stdin on stderr (0.5.8 prints nothing), handoff with no transcript writes nothing and says handoff skipped (no transcript) (0.5.8 records a row that reads asked: n/a | files: none), remember reports boundFromEffect when it read the terms out of the effect slug, and a replacement term under 2 characters is dropped like a withdrawn one. 0.5.8's install-hook --codex also says trust the two memory-pulse entries while writing three hooks: trust all three.

One line, any shell. Installs the Claude Code hooks, the Codex hooks when codex is on PATH, and registers the MCP server. It prints its plan first, runs nothing as root, and downloads nothing itself (npx caches the package under ~/.npm). Append -s -- --dry-run after bash to see the commands without running them, or -s -- --project to write the hooks and the Claude Code MCP registration into the current repository. codex mcp add has no project scope; it writes ~/.codex/config.toml on each machine under either flag.

curl -fsSL https://pulse.strategic-innovations.ai/install.sh | bash

Claude Code, as a plugin. Skill, MCP tools and two hooks (SessionStart brief, PreToolUse guard) from this repo, typed inside Claude Code:

/plugin marketplace add t-crew/memory-pulse/plugin install memory-pulse@memory-pulse

claude plugin details memory-pulse shows what you pay. About 120 tokens are always on, which is the skill's description (Claude Code, 2026-09-01). The four tool schemas resolve at runtime and the hooks are free. The hooks run the plugin's own server.mjs, so what enforces your corrections is the version you installed. The plugin does not carry the PreCompact handoff; to add it, put this entry under hooks.PreCompact in ~/.claude/settings.json:

{ "hooks": [{ "type": "command", "command": "npx -y memory-pulse handoff" }] }

Codex CLI, as a plugin. Same files, read from .codex-plugin/plugin.json:

codex plugin marketplace add t-crew/memory-pulsecodex plugin add memory-pulse@memory-pulse

Then, inside Codex, run /hooks and trust the three memory-pulse entries (session start, prompt and guard). Codex runs no hook it has not shown you, and installing a plugin does not trust its hooks. That is Codex's rule and a good one.

npm, global install. install-hook writes four hooks into ~/.claude/settings.json: SessionStart brief, UserPromptSubmit prompt, PreToolUse guard and PreCompact handoff. --codex targets ~/.codex/hooks.json instead, --project writes them into the repository, and --ambient adds the UserPromptSubmit observer.

npm i -g memory-pulsememory-pulse install-hook

Claude Code or Codex without the plugin. Register the stdio server, then add the hooks:

claude mcp add memory-pulse -- npx -y memory-pulsecodex  mcp add memory-pulse -- npx -y memory-pulsenpx memory-pulse install-hook            # Claude Code: ~/.claude/settings.jsonnpx memory-pulse install-hook --codex    # Codex:       ~/.codex/hooks.json (then /hooks to trust each entry)npx memory-pulse install-hook --cursor   # Cursor:      ~/.cursor/hooks.json (brief, fail-closed guard, capture)

claude mcp add without a scope registers the server for the current directory; -s user registers it once for every project. codex mcp add has no scope flag and writes ~/.codex/config.toml (codex-cli 0.153.2, 2026-09-09). Add --project to install-hook and the hooks are written into the repo, at .claude/settings.json or .codex/hooks.json. Commit that and every clone is re-entered and guarded with nothing for anyone to install. A Codex clone still trusts the entries once via /hooks, and a scripted codex exec needs that trust in place beforehand.

Cursor. Cursor runs hooks, so npx memory-pulse install-hook --cursor (or --cursor --project for .cursor/hooks.json in the repo) gives Cursor the same brief, guard and capture as Claude Code: the brief at session start, the guard on every write and shell command (fail-closed, so a broken guard blocks rather than passes), and capture with fixes and open threads. For the MCP tools, register command: npx with the args below in .cursor/mcp.json or ~/.cursor/mcp.json. Written from Cursor's hook reference (cursor.com/docs/agent/hooks, read 2026-10-21) and tested against its documented payloads; not yet run inside Cursor itself.

Other stdio clients. The same MCP entry works in any client that speaks MCP over stdio.

{ "mcpServers": { "memory-pulse": { "command": "npx", "args": ["-y", "memory-pulse"], "env": { "MEMORY_PULSE_KEY": "" } } } }

A rule file at .cursor/rules/memory-pulse.mdc tells the agent to call pulse before answering about prior work; the guide has the text.

From source. server.mjs plus the plain modules in lib/, zero dependencies:

git clone https://github.com/t-crew/memory-pulse.gitcd memory-pulse && node server.mjs brief --offline

The second command prints the corrections in this repository's own committed ledger, so you see the output format before installing anything.

What each host runs

Read from server.mjs and hooks/hooks.json, 2026-10-21.

First run

The hooks stay silent in a project that has no ledger: brief prints nothing and exits 0. The first remember creates .memory-pulse/events.jsonl, from a tool call or from the shell:

npx -y [email protected] remember pricing-shipped price-corrected --kind correction --withdrawn '$49' --replacement '$29' --note 'measured willingness to pay is $29'# before 0.6.0 is on npm, the same command from a checkout:git clone https://github.com/t-crew/memory-pulse.git && node memory-pulse/server.mjs remember pricing-shipped price-corrected --kind correction --withdrawn '$49' --replacement '$29' --note 'measured willingness to pay is $29'

The hook command install-hook writes is npx -y memory-pulse guard, and npx resolves the package on every call: 0.99 s warm and 1.4 s cold (empty npm cache) per edit on an M1 Pro on 2026-10-16 with 0.8.0, against 52 ms for node server.mjs guard on the same ledger, which is the path the plugin runs. If the cost matters, install the plugin, or install globally and edit the hook command to memory-pulse guard.

Where the key goes

export MEMORY_PULSE_KEY=mp_live_... in the shell that starts Claude Code or Codex; the plugins and the hooks read the shell environment. On claude mcp add, pass --env MEMORY_PULSE_KEY=.... In Cursor, the env block above. Do not also run claude mcp add on a plugin install, or you get two servers.

Corrections are enforced, not only surfaced

Showing an agent a correction is not enough. TRACE (arXiv 2606.13174) measured agents re-violating more than half of the corrections they had been shown, with a memory tool installed. The PreToolUse guard sees every Edit, Write and MultiEdit in Claude Code, plus a Bash command whose write is literal in the command string (a heredoc, a redirect or a tee, read with nothing executed), and every apply_patch in Codex, where one patch may touch several files and each is checked under its own path. An edit that writes back a withdrawn value is blocked, and the agent is told which ledger line retired it and when. An edit that names the replacement beside the old value passes, because "was $49, now $29" is a comparison. Only a bare reintroduction is blocked. A shell command whose write cannot be read from the string returns nothing, so it is a miss rather than a false block, and check --ci on the pull request catches it.

Record corrections with the exact terms:

remember({ cause: "pricing-shipped", effect: "price-corrected", kind: "correction",           note: "measured willingness to pay is $29", withdrawn: ["$49"], replacement: ["$29"] })

The rules that decide whether the guard can act on a correction:

• Matching is a case-sensitive token. A term cannot start inside another word or number. A number cannot run on into more digits but may carry a unit: $49 blocks $49/mo and not $490. A word under five letters cannot run on (ten does not block tenant); a longer one keeps its inflections. n=14 does not block N=14. Record every spelling you have seen, one term each. before matches case-insensitively, and it advises rather than blocks.
• A replacement anywhere in the edit releases every term of that correction. Make replacements as specific as withdrawn terms: $29, not 29. Terms under 2 characters are dropped and the result says enforceable: false.
• Pass the terms yourself. When a correction declares no withdrawn term and its effect reads X-corrected-to-Y (or -withdrawn-for-, -superseded-by-, -replaced-by-), the terms are read out of the slug: price-corrected-to-29 binds the word price. The result then carries boundFromEffect so you can see it happened.
• A wrong correction is fixed with a new one carrying supersedes: [t]. That retires every term of the earlier row, so re-declare what you keep. guard allow is different: it records an override for one path prefix where the old value is legitimate, and the correction keeps binding elsewhere.
• A note that reads like an instruction is refused, including any angle-bracket tag. Write "the script tag" rather than the tag itself. The check reads a note the way a person sees it: invisible characters, letters borrowed from the Cyrillic or Greek alphabets, fullwidth forms and base64 do not hide an instruction from it (since 0.16.0). A paraphrase or a translation can, because the check is a list of patterns, not a reader.

Test the guard by hand with the payload Claude Code sends before a Write (Edit carries new_string, MultiEdit edits[].new_string, Bash command):

printf '%s' '{"hook_event_name":"PreToolUse","tool_name":"Write","tool_input":{"file_path":"docs/pricing.md","content":"The price is $49."}}' | npx memory-pulse guard; echo "exit $?"

Exit 2 with the ledger line means the guard is live. Stdin that is not hook JSON is allowed through silently; guard --verbose says so on stderr, which is how you tell a miswired hook from "no evidence".

Commands

npx memory-pulse help           # every command and environment variable; also --help, -hnpx memory-pulse serve          # the MCP server on stdio (the default with no command)npx memory-pulse brief          # the SessionStart re-entry: the organized memory (scaffold) by default; with a key, --engine or --budget the engine briefnpx memory-pulse scaffold       # the organized memory: identity and rules, corrections in force, project state, recent findings, open threads; --section S, --allnpx memory-pulse telemetry show # the opt-in usage counts: exactly what would be sent once a day (on, off, reset)npx memory-pulse prompt         # UserPromptSubmit hook: the corrections and decisions that bear on the prompt, injected as contextnpx memory-pulse brief --budget 2000   # the richest tier that fits 2000 tokens, corrections first and wholenpx memory-pulse brief --offline  # the plain local list; an engine error falls back to the organized memory insteadnpx memory-pulse remember <cause> <effect> [--kind correction] [--withdrawn T]... [--replacement T]... [--supersedes N] [--note "..."] [--tags a] [--pinned] [--scope agent]npx memory-pulse before "<what you are about to do>"   # which corrections bear on the change; exact rung local, learned rung from the enginenpx memory-pulse guard          # PreToolUse hook: blocks edits that reintroduce withdrawn terms (exit 2); --verbose names a non-JSON stdinnpx memory-pulse guard allow "<term>" --path <prefix> "<reason>"   # record a false block as an overridenpx memory-pulse check --ci     # Memory CI: one of three verdicts for a change, from files you ownnpx memory-pulse verify         # row chain, last engine seal, this machine's anchor and writer marks; on a break: the first bad row and the command that fixes it; exit 2 if any fails; --strict also fails on unmarked rows; --json for machinesnpx memory-pulse append --json '{"cause":"...","effect":"..."}'   # the one way for another tool to write a row: lock, t, prev and hash done here; also `-` for stdin, --scope agentnpx memory-pulse repair --quarantine-tail   # move a tail of rows that fail on their own aside, byte-for-byte, recorded; nothing deleted (also unchained rows this machine did not write)npx memory-pulse repair --adopt-quarantine <file> --yes   # re-append quarantined rows through the writer after a person has reviewed themnpx memory-pulse seal accept "<why>"   # a stale seal over rows that verify: record the decision, set the old seal aside; refused on a broken chainnpx memory-pulse anchor         # how far this machine's anchor reaches into the ledger, and which rows its writer did not write; --scope agent, --jsonnpx memory-pulse anchor accept "<why>"   # after a restore, or rows from another machine or an older client: record that you trust the ledger as it standsnpx memory-pulse state          # where this ledger stands: rows, in force, rolled back, heads, seal, branchnpx memory-pulse state export > agent.rain   # the .rain state capsule: rows verbatim, chain head, set head, seal, invariants, one digestnpx memory-pulse state import agent.rain     # on another machine: verify, install byte-identical (or fast-forward), record where it came fromnpx memory-pulse state branch exp --at 12 --out exp.rain   # a capsule cut at row 12, named exp; the source ledger is untouchednpx memory-pulse state merge exp.rain        # carry a diverged branch's findings across once; conflicts are named, never resolved quietlynpx memory-pulse state rollback --to 12 "<why>"   # one appended row; rows after 12 stay on record, out of forcenpx memory-pulse state verify agent.rain     # verify a capsule without touching any ledger; exit 2 on failurenpx memory-pulse board post --from blue --to science --kind request --title "audit stage 12" --body "..."   # the shared agent message boardnpx memory-pulse board read --as science      # unread messages for an agent; also reply, ack, thread, html (the board view), verifynpx memory-pulse board verdict 4 refuted --as coder --defect "..."   # answer a claim; a refutation names its defect; claims lists every claim and its standingnpx memory-pulse board sync on --project <id>   # push the board after every write; the console shows each agent's open asks and every message (board push / status by hand)npx memory-pulse login          # connect this machine to a hosted project: approve it in the console and the key is saved for you, no copying; turns board sync onnpx memory-pulse logout         # remove the saved key from this machine (revoke it in the console to end it everywhere)npx memory-pulse capture        # PostToolUse hook (installed by install-hook): one observation per tool call into the session's episode, local, no modelnpx memory-pulse consolidate    # Stop / SessionEnd hook: fail-then-pass becomes a fix and still-failing an open thread, each pointing at its evidencenpx memory-pulse weights        # what each memory did: check outcomes shown vs. sat out at random; retired ones markednpx memory-pulse adopt          # the corrections you already keep (Claude Code feedback memories, CLAUDE.md, AGENTS.md, rules files), shown; --yes enforces themnpx memory-pulse lint [--ci]    # do CLAUDE.md / AGENTS.md / .claude/rules still state a value the ledger retired?npx memory-pulse report         # correction re-violation scoreboard, computed locallynpx memory-pulse bench          # instant measured metrics on YOUR ledgernpx memory-pulse stats          # your telemetry capsule, signature verified by the enginenpx memory-pulse badge          # README badge markdown from your own signed numbersnpx memory-pulse install-hook   # four hooks (idempotent); --codex targets Codex; --project commits them; --ambient adds the observernpx memory-pulse mode [agent|soul|deliberate]   # show or set the modenpx memory-pulse identity "<name>, <role>; <stance>"   # pin the agent's self on the agent ledgernpx memory-pulse soul "<persona, voice, values>"   # soul mode: the agent's soul; confirm / reject held changes; history / revert; import / export --openclawnpx memory-pulse user "<who you are>"   # soul mode: the agent's notes about the person it works withnpx memory-pulse taint          # PreToolUse hook that mode soul installs: marks a session that reads outside contentnpx memory-pulse lesson         # the lesson bank the identity block prints, with where each lesson held; --held-out lists what was left out and whynpx memory-pulse lesson probe 41 held --where billing-api   # record that a carried lesson held (or failed) there: the evidence the bank ranks bynpx memory-pulse lesson freeze  # pin the bank you reviewed; every brief prints it until the next freeze or `lesson thaw`npx memory-pulse handoff        # PreCompact hook: record what the session was doing, from the transcript on stdinnpx memory-pulse ambient        # Stop hook (agent mode): decision, preference, lesson or correction from the last turnnpx memory-pulse observe        # UserPromptSubmit hook: a prompt shaped like a correction is recorded with both terms

Undo: /plugin uninstall memory-pulse inside Claude Code; delete the memory-pulse entries under hooks in ~/.claude/settings.json or ~/.codex/hooks.json; claude mcp remove memory-pulse; delete .memory-pulse/ in a repository; delete ~/.memory-pulse/agent/. Each step leaves the ledgers in place except the last two.

The plugin also ships a skill at skills/memory-pulse/SKILL.md that teaches the agent when to pulse, how to record corrections with withdrawn terms, and how to respect the guard. The npm tarball ships server.mjs, python/, this file and the licence; the skill, the hooks and the Action live in the repository.

What the brief tells you before it tells you anything

Every engine-backed brief opens with one line of provenance, so a session can tell whether its memory loaded whole, truncated, or not at all:

memory-pulse: loaded 852 events from .memory-pulse/events.jsonl · sha256 1a2b3c4d5e6f · 2 binding corrections (10 withdrawn terms) · 1 superseded · ⚠ 1 malformed line skipped: 544 · memory key resumed (+3 new) · tier brief, 5,153 chars

The local render (brief --offline, or the engine unreachable) opens with memory-pulse: engine unreachable — local render from N events and carries no sha256 and no binding count; lint prints the provenance line offline.

Every CORRECTIONS line cites its ledger record as … -> effect (t12) — note, so a correction is evidence the agent can point at. recall and the guard name the same t. The engine's brief cites each correction with its note and prints no withdrawn terms, and it lists superseded rows; the offline render prints withdrawn: a, b → replacement and filters superseded rows.

Tiers, from the engine's src/disclose.js and src/tokens.js read 2026-09-09: index 1,800 chars, brief 6,000, notes 20,000, full 120,000. With --budget N the richest tier whose ceiling fits N tokens at 3.8 chars each is chosen, so brief needs 1,579 tokens, notes 5,264 and full 31,579; --budget 1500 gives the index tier. MEMORY_PULSE_BRIEF_BUDGET and MEMORY_PULSE_BRIEF_TIER set the same for the hook. The brief costs context: 5,153 chars, about 1,350 tokens, on the 852-event ledger above, 2026-09-01.

An AT RISK block appears after the guard has blocked the same correction twice; before "<what you are about to do>" names the corrections that bear on a change before you make it.

Lint: the rules a session loads, checked against the ledger

Governance files drift. A CLAUDE.md written in June still says the price is $49 after the ledger retired it in August, and every new session loads the stale rule with full confidence. lint runs the guard's check over the files a session will read and gives each of them one of the three verdicts. It covers CLAUDE.md, AGENTS.md, .claude/CLAUDE.md, .claude/rules/, .cursorrules, .cursor/rules/, .github/copilot-instructions.md, .codex/AGENTS.md, Muse Code's .agents/AGENTS.md and .agents/memory/, .memory-pulse/invariants.jsonl, and any paths you pass:

$ npx memory-pulse lintmemory-pulse: loaded 2 events from .memory-pulse/events.jsonl · sha256 8e401a39f323 · 1 binding correction (1 withdrawn term)  BLOCKED     CLAUDE.md             • "$49" was withdrawn at ledger t2: price-49-launched -> price-corrected-to-29 — use $29  verified    AGENTS.md  no evidence .claude/rules/style.mdlint: 3 file(s) — 1 blocked, 1 verified, 1 no evidence — a rule your ledger retired is still being loaded into sessions

Exit 2 on any blocked file. Under --ci it also exits 1 when it found nothing to check, so an empty run is never reported as a pass. --json is there for machines. It also names the corrections that carry no withdrawn terms, which surface in the brief but which nothing can enforce.

A line in any of those files that reads like it was planted to hijack the agent (an order to ignore its previous instructions, to send keys or tokens to an address, to keep something from the user, or a <system> tag) is reported as a WARNING with the file, the line number and the pattern; the line itself is never printed. Lint still exits 0 on a warning, because no pattern can tell every API doc ("send the token to https://auth...") from a planted line; lint --strict blocks on it. This matters most for committed agent memory: Muse Code loads a repository's .agents/memory/ even in a workspace you have not trusted, so run lint --strict on a checkout you did not write before an agent opens it. On 282 memory and instruction files and 1,500 package READMEs (about 278,000 lines) it warned on nothing.

Tamper evidence: the ledger cannot be edited quietly

Since 0.3.1 three mechanisms hold, each with its own job:

• Row chain. Every row remember writes carries prev, the previous chained row's hash, and hash, the SHA-256 of its own canonical JSON. Rows that existed before the chain are never rewritten, and the first chained row seals them with a digest. An edit in place, a removed row, a reordered row or an unchained row after the chain started all fail verifyChain(). A failed chain blocks every check and lint verdict, because a memory whose own history is in question cannot vouch for anything. The guard runs on every tool call, so since 0.6.0 it is narrower: a break in the middle of the file blocks writes to the ledger and its sidecars and warns on everything else, and a bad tail is moved aside (below). Reading is never blocked.
• Set head. The engine also commits to the ledger as an order-free fold, a multiplicative group mod a 3072-bit prime, which is the MuHash construction Bitcoin Core uses for its UTXO set. Shards from several agents fold to the same head in any order, and removing a row is the group inverse, so the state stays exact while the history stays append-only. A literal XOR fold was measured forgeable, since a linear system hides an edit in 10 ms at 300 rows, and it is not used here.
• Seal. Every read call returns a seal signed by the engine, carrying the row count, the watermark, the set head over every row's full content, and the chain head. The client keeps it in seal.rain beside the ledger and presents it on the next call. Locally, rows up to the sealed watermark must fold to the sealed head before the ledger is trusted. At the engine the signature and the fold are re-checked, and an edit below the watermark is reported as drift and blocks. A process with write access can rewrite the file and even the chain. It cannot produce the engine's signature, and it cannot make edited rows fold to the sealed head.

Nothing is stored server-side for any of this. The seal travels in the payload, the same way the telemetry capsule and the memory key do.

An earlier copy put back is refused (0.16.0). The seal and the chain check the ledger you show them. Put back an earlier copy, its seal with it, and both agree with it, because every row in it once was the record. So each machine now keeps an anchor for each ledger it writes or checks: how many rows it last saw and a digest of exactly those rows. The anchor lives outside the ledger's folder, in ~/.local/state/memory-pulse/ (or MEMORY_PULSE_TRUST_DIR), and it only moves forward. When a ledger no longer begins with the anchored rows, verify exits 2 and says so, and the guard blocks writes to the ledger, its files and the anchor itself. Other work goes on. Restore the newer copy, or run memory-pulse anchor accept "<why>", which records the decision as a row.

Every row says whether this machine wrote it (0.16.0). The writer marks each row it appends with a key that sits beside the anchor: 32 random bytes, readable by you only. The mark is inside the hashed row, so older versions of memory-pulse still verify the chain. From the moment the anchor starts:

• a marked row whose content changed is named, and blocks writes to the ledger like a broken chain. Re-chaining the file does not hide it, because re-chaining needs no key and marking does;
• an unchained row at the end of the file is moved aside by the next write or guard run, recorded like a bad tail, instead of being sealed into the chain. No memory-pulse since 0.3.1 writes an unchained row;
• any other row without this machine's mark is named by verify and accepted, and the anchor covers it from the next write on: a row from another machine, from a memory-pulse older than 0.16.0 still installed somewhere, or one someone chained into the file by hand. Nothing on the row tells these apart, and the first two are everyday traffic while a machine upgrades. Where one writer is all there is, verify --strict (or MEMORY_PULSE_STRICT_WRITER=1) fails on them, and memory-pulse anchor accept "<why>" records that you trust them.

What the anchor and the marks stop, and what they do not. They stop a process that can write the ledger's folder but not the trust store, which is a different folder: a sandboxed tool, a sync or backup that restores an old file, git checkout of an old ledger, a commit that rewinds .memory-pulse/. They do not stop a process running as you, which can rewrite both folders. Against that, the engine's signature still catches an edit below the seal's watermark on the next engine call. A whole earlier state put back together with its seal and its anchor passes everything on the machine; catching it needs a record somewhere that process cannot write, such as another machine, and 0.16.0 does not keep one. The anchor starts the first time 0.16.0 sees a ledger, so it trusts whatever is there at that moment. A row written through memory-pulse itself carries a valid mark whatever it says: what a row says is the instruction check's job, and that check matches patterns in the text. It does not judge what a row means.

Never write a ledger file directly. Another tool that wants to record a row runs memory-pulse append, which takes the ledger lock (events.jsonl.lock) and computes t, prev and hash itself; two writers never share a t. A row written by hand does not chain: on 2026-10-02 a script hashed a row over Python's default json.dumps separators and reused the previous t, and every session on the machine was refused every edit until a person removed it.

A bad tail is moved aside, not fatal. When every row after the last good one fails on its own (unparseable, its own hash wrong, a t that does not increase, a link to nothing), the guard and the writer move those rows byte-for-byte to events.jsonl.quarantine/<time>.jsonl, append one chained row recording the file, the row count, the SHA-256 of the moved bytes and the reason, print one warning line, and let the work continue. Nothing is deleted. memory-pulse repair --adopt-quarantine <file> shows what would be re-appended; with --yes, after a person has read them, the rows go back through the writer with new t, prev and hash and an origin pointing at the quarantine file.

A break in the middle is restored, not repaired. A hand edit with valid rows after it blocks writes to the ledger and its sidecars until the file is restored from a copy you trust; verify names the first broken row and lists the copies beside the ledger that verify. A stale seal is a different case: the rows verify, but the engine's last seal no longer matches them, because something rewrote and re-chained the ledger or the engine sealed a different view. memory-pulse seal accept "<why>" records that decision as a ledger row, sets the old seal aside beside the ledger, and the next engine call issues a fresh one. Neither case is fixed by deleting the ledger, and the guard's message never suggests it.

What to commit. events.jsonl, invariants.jsonl and, if one exists, events.jsonl.quarantine/ (the record of what was moved aside). Keep seal.rain, seal.resume.json, verify.cache.json, events.jsonl.lock, telemetry.rain, memory.rain, violations.jsonl and episodes/ out of git; this repository's .gitignore carries that list, and nothing writes a .gitignore for you below 500 events. The anchor and the writer key are not in the repository: they are one machine's view, and a fresh clone starts its own anchor the first time it runs. The ledger is read from the directory the agent started in, so start at the repository root or set MEMORY_PULSE_LEDGER.

Agent state travels as a .rain capsule

An agent's state is its ledger, so moving the agent means moving the ledger with its integrity intact. memory-pulse state export packs the ledger as a .rain state capsule: every row verbatim, the row-chain head, the order-free set head, the engine's last seal and the declared invariants, under one digest. On another machine, memory-pulse state import verifies all of that before it trusts the capsule, installs the rows byte-identical (or fast-forwards a ledger the capsule extends), and records where the state came from as a row. The corrections travel with it: the guard blocks on the new machine exactly as it did on the old one. A capsule with one edited row is refused before anything is written.

Three more operations, none of which rewrites a row:

• state branch <id> --at <t> cuts a capsule at row t and names it. Import it elsewhere and work on it; the source ledger is untouched.
• state rollback --to <t> "<why>" appends one row. Every row after t stays on the record and goes out of force: a rolled-back correction stops binding, a rolled-back lesson leaves the identity block, and a later rollback to a higher row restores them. brief, guard, check, lint, before and report read the in-force view; the engine reads the same view and seals the full record.
• state merge <file> carries a diverged branch's findings across once, in the branch's order, each carrying its origin. Duplicates are skipped by content, the branch's own rollbacks are applied at source, a note that reads like an instruction is held back, and a correction that withdraws a term this ledger withdraws with a different replacement is named as a conflict, not resolved. Merging the same capsule twice appends nothing, and two ledgers that merged each other in either order hold the same findings in force.

state verify <file> checks a capsule without touching any ledger, and memory-pulse state alone says where the ledger stands. --scope agent does all of this for the agent ledger. The seal's engine signature is checked by the engine on the next call; locally the capsule's rows must fold to the sealed head. No server is in the loop for any of it.

Agent message board: shared memory between agents

Several agents (a coding agent, a research agent, a reviewer) can share one board: a ledger file of its own that every one of them can post to and read. Point each agent at the same path with MEMORY_PULSE_BOARD (a shared folder works) and they talk through it:

npx memory-pulse board post --from science --to blue --kind request --title "need a forward dump" --body "predicted peaks for the near-miss pairs"npx memory-pulse board read --as blue        # t1 · science → blue · request · need a forward dumpnpx memory-pulse board reply 1 --as blue --body "dump written to from-science/fwd"npx memory-pulse board ack 1 --as blue       # out of blue's unread list; kept on the boardnpx memory-pulse board html --out board.html # one offline page: who writes to whom, every thread, the chain status

The board is chained and locked like every memory-pulse ledger, so an edited message fails board verify and every reader sees the break. It is passive: posting runs nothing, and nothing here wakes, schedules or prompts an agent; each agent reads the board when it next checks. A message is data from another agent, never an instruction: text that reads like one is stored, flagged and shown quoted, and the board never enters the project ledger, the brief or the guard. With MEMORY_PULSE_BOARD set, the MCP server adds two tools, board_post and board_read, so an agent can use the board without a shell.

A team that agrees early can drop the one member who was right. So a claim (a position with its support) stays open until another agent posts a verdict on it: upheld, refuted, or withdrawn by its author. A refutation must name the step, number or fact that fails, and no agent rules on its own claim. A synthesis is refused while a claim in its thread is open; --over-open records it anyway, flagged with the claims it went over.

npx memory-pulse board post --from science --kind claim --title "hit rate under 20 percent" --body "last week's trace" --thread 1npx memory-pulse board verdict 2 refuted --as coder --defect "the trace excluded the batch jobs that dominate reads"npx memory-pulse board claims                # every claim: open, upheld, refuted, contested or withdrawn, with each verdict

This follows Self-Organizing Agent Teams (arXiv 2609.22682). Six of the twenty strategies its teams kept exist to keep a claim the consensus dropped in play, and its judge may reject an answer only by naming a specific defect.

It remembers what the agent did, on its own. The plugin, install-hook and install-hook --cursor (not yet --codex) add a capture hook after every tool call: an edit (file, lines changed), a command (scrubbed of secrets, with its outcome and first error line) or a read becomes one row in the session's episode, .memory-pulse/episodes/<session>.jsonl. No model runs and nothing leaves the machine; tool output is never stored, only its outcome. At the end of each turn the episode is consolidated: a check that failed and later passed becomes a fix ("npm test failed: …, passed after editing src/price.js"). A check still failing when the session ends becomes an open thread, closed by a later fix or a passing run of the same check. Each finding names the episode rows it rests on and the chain head after them, and at the end of the session the episode is saved as a .rain capsule, so the evidence behind any finding can be checked without a key. The prompt hook brings a past fix back when you mention the same file. --no-capture leaves it out.

Connect a machine with npx memory-pulse login. It prints a short code and opens the console; sign in, check the code, pick the project and approve. The console shows where and when the request was made and warns when that is not where you are, so deny anything you did not just start. The key is saved to ~/.config/memory-pulse/key (only you can read it) and every command uses it when MEMORY_PULSE_KEY is unset, sent only to the engine it came from; if it is revoked, commands fall back to the free tier and say so. Login also turns board sync on for the board in the current folder and pushes it. Logging in again replaces the key this machine was using (the CLI sends a hash of that key to show it holds it), so keys do not pile up; a login from a machine without one adds a key. npx memory-pulse logout removes the saved key; revoke it in the console to end it everywhere.

See your agents' board in the console. With hosted memory on for a project (Starter, Pro, Enterprise), board push sends the board there with the project's key, and board sync on --project <id> pushes after every post, reply, ack and verdict, from the CLI and from the MCP tools. Sign in at the console, open the project's Board: each agent's open asks, what it is waiting on, its open claims and when it was last active, then the newest 200 messages, shown as text. After the first push only new rows travel. A failed sync never fails a post; it says so and board push sends it later. The hosted board only fast-forwards, because replies and acks name rows by number, so one board per project. Turning hosted memory off erases the hosted board with it.

MEMORY_PULSE_KEY=<project key> npx memory-pulse board sync on --project <id>npx memory-pulse board status                # local rows, sync state, and the hosted board's count

Survives compaction, works offline, captures corrections, speaks Python

Four things added on 2026-09-03, each deterministic (no model in the loop):

• Compaction handoff. install-hook adds a PreCompact hook (the plugin's hooks.json does not; see Install). Before Claude Code compacts, memory-pulse handoff reads the transcript and records what the session was doing as facts: the last asks, the files edited, the last error, and the assistant's last state. The next session start prints it first, online or offline. An instruction-like message is dropped from the note and never recorded; a call with no transcript writes nothing.

• Offline brief. When the engine is unreachable, whether air-gapped, on a dead network or during an outage, the session no longer starts empty. brief prints a local render carrying every binding correction with its withdrawn and replacement terms, the last handoff and the recent rows, labelled as a local render with no ranking applied. brief --offline forces it. Guard, check, lint and verify never needed the network. The MCP pulse tool returns an error instead, so a client with no hook starts empty when the engine is down.

• Ambient correction capture, opt-in. install-hook --ambient adds a UserPromptSubmit hook. A prompt shaped like a correction, such as the price is $29 not $49, change 0.3.1 to 0.3.2 or 500 events -> 924 events, is recorded as a correction carrying both terms, so the guard enforces it from the next edit on. A prompt that does not yield both terms is left alone. Silent unless --verbose.

• Python client. python/memory_pulse.py is a single stdlib-only file with the same ledger format, the same hash chain and the same guard rule. A LangChain or CrewAI agent and a Claude Code session can share one ledger and verify each other's rows. The test suite writes rows from Python and verifies them in Node, and back.

• Decode-time guard for local models. python/span_guard.py applies the same rule one layer down. The token that would complete a withdrawn value is masked while the model decodes, so the value cannot be generated, and the replacement's next token is offered in its place. It reads the ledger under the rules check uses, so a superseded correction is not enforced and a broken chain refuses to build. Stdlib only, and its suite runs without a model. Measured with mlx-lm on TinyLlama-1.1B and Qwen2.5-1.5B on 2026-09-03, the same prompt whose plain decode wrote the withdrawn price wrote the corrected one under the guard, and the sentence around it stayed intact. It blocks the spellings it was given, including the written variants you record, and a paraphrase nobody enumerated gets through.

  python

  from memory_pulse import Ledgerfrom span_guard import SpanGuardguard = SpanGuard.from_ledger(tok, Ledger(".memory-pulse/events.jsonl"))# mlx-lm: generate_step(..., logits_processors=[guard.logits_processor(mx, np)])

Soul mode: an autonomous agent's own identity, voice and memory (opt-in)

Soul mode is agent mode plus a soul: the persona, voice and values the agent keeps, its notes about the person it works with, its long-term memory and a journal of recent days. The agent owns them. It can rewrite its soul, its notes and its name from any session, and every brief in every tool starts with the current version and the agent ledger's fingerprint.

npx memory-pulse mode soulnpx memory-pulse soul "Warm and direct. Speaks plainly, prefers measured numbers to adjectives, says so when unsure."npx memory-pulse user "Travis, founder; wants the reasoning recorded, not just the result."

An agent that can rewrite itself can be talked into it. OpenClaw keeps the same things in workspace files, and the Cloud Security Alliance documented the attack: a poisoned page or mail plants instructions in SOUL.md and MEMORY.md that survive restarts. Here what decides is where a change comes from. A hook the host runs before WebFetch, WebSearch and every MCP tool but memory-pulse's own (mode soul installs it), and the guard for curl or wget, mark the session as having read outside content, keyed by the session id the host supplies, not one a command can change. A change made in such a session is held, and takes effect only when the agent confirms it from a session that did not (soul confirm), or you approve it; from a marked session the guard refuses confirms, reverts and agent-ledger state operations. Soul text that reads like an injected instruction, tells the agent to rewrite its own files, or tells it to obey what arrives from outside is refused, with the reason. Every version is kept with its session and date: soul history lists them and soul revert brings one back. The agent writes its log through memory-pulse, so a raw edit of the ledger file is blocked and every entry stays chained. This needs the guard (the plugin or install-hook) and runs on Claude Code today. soul import --openclaw reads an OpenClaw workspace through the same screens and names every line it withheld; soul export --openclaw <folder> writes the files back, never over a file it did not write, and the brief says when one of them was edited afterwards.

What this does not stop, stated plainly: an agent that starts a new session itself (claude -p) and confirms there; outside content read through a channel no hook sees (gh, git pull, a downloaded file, a subagent, a script that fetches); a command hidden in a script file, which the guard cannot read; a host other than Claude Code; wording the screens do not know, since they catch known shapes only; and a correctly chained row written by code that runs as you. What holds instead is that a change shows: a row appended around memory-pulse, or one with a made-up hash, is not honored as the soul and is named in the brief, and the fingerprint (the sha256 of the agent ledger's bytes) changes with any edit. A soul or an open change arriving by state merge is held, and once the agent ledger holds a soul these rules apply whatever MEMORY_PULSE_MODE says. mode soul warns when the guard or capture hook is missing.

Agent mode: a persistent agent identity that grows (opt-in)

The default is deliberate. Memory lives in the repo and capture is explicit. Agent mode covers the other thing people ask for, which is an agent that is the same agent tomorrow, in every project and every tool, and that grows.

npx memory-pulse mode agentnpx memory-pulse identity "Blue, research agent for Travis; innovate, don't debate"

What that turns on:

• An agent ledger at ~/.memory-pulse/agent/events.jsonl, belonging to the agent and shared by every project and every tool that speaks MCP. Same format, same hash chain, same seal.
• A self block, first in every brief, online or offline. It carries who the agent is as a pinned line, the standing rules and preferences it has learned (rows tagged rule and preference), a bank of at most ten lessons (tagged lesson, chosen as described below), and a fingerprint made of the chain head and the engine's seal, so the agent can state which memory it is running on and show it was not swapped or edited overnight.
• Growth, after every turn. Two user-scope hooks in ~/.claude/settings.json (Claude Code only, today). The Stop hook deterministically records a stated decision to the project ledger, a user preference or stated lesson to the agent ledger, and any correction shaped like one. It is capped at four rows a turn, tagged ambient, never pinned, and never taken from instruction-like text. Identity itself is only ever set by you or superseded by a correction.
• Corrections that follow the agent. A correction on the agent ledger blocks the same edit in any project, in either mode, for as long as the file exists.
• Lessons kept for what they did, not for being new. The bank holds pinned lessons first, then lessons that held somewhere other than where they were learned (chosen to cover as many places as possible), then two new lessons on probation. A lesson that failed as often as it held leaves the bank. A sentence that only points at a lesson ("the lesson is in the ledger"), a fragment and a duplicate are held out and counted. memory-pulse lesson probe 41 held --where billing-api records the evidence; memory-pulse lesson freeze pins the bank you reviewed. The method follows Self-Organizing Agent Teams (Pappu et al., arXiv 2609.22682), where a team keeps a learned strategy only after it works on problems other than the one that produced it.

npx memory-pulse mode deliberate turns the hooks off again and leaves the ledgers in place. remember takes scope: "agent" from any tool and --scope agent from the shell, in either mode. The rule of thumb: off for one repository; on when one agent spans repositories and tools and you accept up to four rows per turn, with the decision rows landing in the project ledger you commit.

What runs where (the privacy contract)

• Your ledger is a local file at .memory-pulse/events.jsonl in your project. You can commit it, grep it or delete it.
• remember writes to it directly and works offline.
• Read operations send the ledger's events to the hosted engine over TLS, which computes the answer and forgets the request. Unless you turn on hosted memory for a project (opt-in, in the console), the service keeps no copy of your memory. State arrives in the request and leaves in the response. Hosted memory and the hosted board are stored until you turn them off, which erases them.
• Telemetry is a signed capsule beside your ledger at .memory-pulse/telemetry.rain. The engine advances it on each read call and hands it back without ever storing it. stats verifies the signature and badge turns it into a README badge. Delete the file and it restarts.
• State persistence without storing your ledger. After a read the engine hands back a signed memory key (.memory-pulse/memory.rain, git-ignored). The next read presents it and the engine resumes from it, ingesting only the events recorded since. The answer is byte-identical to a full rebuild. Any mismatch, whether edited history, a stepped ledger size or a bad signature, falls back to a rebuild and says why. Lose the file and you lose nothing but one rebuild. MEMORY_PULSE_MEMORY_KEY=off disables it.
• Memory integrity. A note that reads like an instruction, such as "ignore previous instructions", "run this command" or a fake system tag, is refused by remember. One already sitting in a ledger is quarantined at read time and reported, so memory never reaches your agent's context as an order. Invisible characters, look-alike letters and base64 do not hide one; a paraphrase or another language can. The signed capsule also raises a drift alert when a ledger loses corrections, shrinks, or its usage shape jumps, and the brief footer shows it. Both checks are deterministic lists you can read.
• This client is the entire client: server.mjs and three plain modules in lib/, zero dependencies, readable in one sitting.

Usage counts

Usage counts are off by default. Nothing is counted until you say yes: install-hook asks once on a terminal, the one-line installer asks through the terminal, and memory-pulse telemetry on turns them on at any time. Neither ever asks in CI. When they are on, the client sends at most once a day, to /v1/usage:

No ledger text, path, term or prompt is in it, and none can be: the client accepts only those eight names. memory-pulse telemetry show prints exactly what would be sent. memory-pulse telemetry off, MEMORY_PULSE_TELEMETRY=0 or DO_NOT_TRACK=1 stops it.

Environment

Pricing

• Free. Ledgers up to 500 events, 200 engine reads per IP address per day (worker/api.js, read 2026-09-09). A session start is one read; past the limit the brief falls back to the offline render and says why. No account, no key.
• Starter, $7/mo. Ledgers to 5,000 events, 1,000 hosted reads a day, the hosted memory (ranked brief, recall, code mode) and the memory-pulse skill.
• Pro, $19/mo. Ledgers to 20,000 events, unlimited reads, signed check receipts. One environment variable, MEMORY_PULSE_KEY.
• Enterprise, $79/mo. Ledgers to 250,000 events, ten revocable seat keys on one subscription, ledger receipts through /v1/certify that anyone verifies keyless at /v1/verify, 14-day refund.

Local writes are free on every tier.

Measured, on our own ledger

Measured on the ledger of the project that builds memory-pulse, a 767-event file of 1.08 MB, measured 2026-09-01:

• A cross-referencing question answered through execute returned 124 chars against the 1,080,983-char full dump. The intermediates never entered context.
• Re-entry briefs at the smallest tier run ~99% smaller than reading the ledger in.
• On our recall benchmark of 351 distinct causes, the noise-floor gate returned zero wrong top answers. When it could not clear the floor it returned nothing.

The ratios depend on ledger size, and a ten-event ledger has nothing to compress. The methodology lives in the engine's benchmark suite.

FAQ

Why is the engine hosted? The ranking engine is the part that took the research. We chose a local ledger, a thin auditable client and a hosted engine over shipping a weaker local ranker. If a remote engine is a dealbreaker, MEMORY_PULSE_API points the client anywhere.

What about team memory? Commit .memory-pulse/ to the repo and your teammates' agents pulse the same ledger. Rows chain on t and prev, so two branches that both append conflict on merge: never keep both sides, re-record the losing rows with remember, and run verify. Shared hosted ledgers are on the roadmap.

License? Client: MIT. Engine: proprietary, hosted.

MIT © Travis Crew

Memory CI: three verdicts, including an explicit empty one

memory-pulse check gives a change one of three verdicts, computed locally from your ledger and your declared invariants:

• blocked. The text writes back a value a correction withdrew, and the verdict names the ledger line that retired it and what to use instead. Also fires when the text trips a declared invariant. Exit 2.
• verified. Recorded events bear on the text and none is contradicted. Exit 0.
• no evidence. The ledger has nothing to say, reported as exactly that and never as a pass. Exit 1 under --ci. The hook stays silent on this verdict, because a hook that comments on every edit livelocks the agent, and CI is where it is loud.

npx memory-pulse check --ci --diff origin/main      # added lines of the branchnpx memory-pulse check --ci --file docs/pricing.mdecho "price is $49" | npx memory-pulse check --cinpx memory-pulse check --receipt --text "…"        # engine-signed receipt, keyless verify at /v1/verify (Pro, Enterprise)

Invariants are declared, never inferred. They live one per line in .memory-pulse/invariants.jsonl, shaped like {"id":"receipt-wording","statement":"say tamper-evident", "patterns":["/\\bproof\\b/i"],"paths":["site/"],"severity":"block"}. A pattern written /…/flags is a regular expression and anything else is a verbatim substring. paths scopes the rule to path prefixes, so a rule about public wording does not fire on a test fixture. severity: "warn" reports without blocking.

Overrides are the false-block signal. guard allow "$49" --path docs/history "historical table" records an override event scoped to that path prefix. The hit passes there and nowhere else, and both report and the signed capsule count it. The guard never guesses, and only explicit withdrawn terms and declared invariants can block. Measured on our own 852-event ledger with the guard-eval bench in the engine repo on 2026-09-01: precision 1.0, zero false blocks over 871 negatives including 694 real notes, and p95 of 1.7 ms per in-process check at 1k events.

PR status check with uses: t-crew/memory-pulse@v0

The same three verdicts as a GitHub check on every pull request, against the ledger and invariants committed in your repository:

# .github/workflows/memory-ci.ymlon: pull_requestpermissions: { contents: read, checks: write, pull-requests: write }jobs:  memory-ci:    runs-on: ubuntu-latest    steps:      - uses: actions/checkout@v5        with: { fetch-depth: 0 }      - uses: actions/setup-node@v6        with: { node-version: 22 }      - uses: t-crew/memory-pulse@v0        id: memory        with:          ledger: .memory-pulse/events.jsonl   # default          lint: "true"                         # also lint CLAUDE.md, AGENTS.md, .cursorrules and the rest          api-key: ${{ secrets.MEMORY_PULSE_KEY }}      - run: echo "${{ steps.memory.outputs.verdict }} / ${{ steps.memory.outputs.conclusion }}"

It checks the added lines of the pull request. blocked fails the check, and the sticky comment cites the ledger line that retired each value. verified succeeds with "N memories checked, M corrections enforced", where M is the ledger-wide count of binding corrections. no_evidence is a neutral conclusion with an explanation, so an empty evidence set is never reported as a pass. The comment is one per pull request and updated in place. This repository runs it on itself, which is why .memory-pulse/events.jsonl is committed, and the first pull request it blocked is the demo.

Details that bite:

• fetch-depth: 0 is required; without it the diff falls back to the working tree, the job log says could not compute the diff ... checking the working tree instead, and the comment reports no evidence on an empty diff.
• base is a bare branch name (main), defaulting to the pull request's base.
• The added lines are checked under the path pull-request, so path-scoped overrides and path-scoped invariants do not fire in CI. Record an override without --path on the pull-request branch when a historical value is legitimate.
• Two status items appear, the job and a check-run named memory-pulse. Require the check-run in branch protection.
• Forks get log-only verdicts: no token, no comment.
• lint runs by default and blocks on a stale CLAUDE.md; lint: "false" turns it off.
• api-key is exported to the CLI, but the Action does not request a receipt today. For a signed receipt run npx memory-pulse check --ci --diff origin/main --receipt as a separate step with MEMORY_PULSE_KEY set.

Releasing

scripts/release.sh 0.6.0

The script gates on fail 0 from the test suite, bumps the four version files together, opens and merges the release pull request, tags, and deploys the engine. The tag triggers .github/workflows/release.yml, which runs the tests and publishes to npm and the MCP registry using GitHub's OIDC identity. There are no tokens in the repo and no one-time passwords. npm attaches provenance automatically, so anyone can verify the package was built from this repo.

One version has four readers: package.json for npm, server.json for the MCP registry, .claude-plugin/plugin.json for Claude Code and .codex-plugin/plugin.json for Codex. All four must agree. test/manifests.test.js fails until they do, and the release workflow refuses to publish while they disagree.