memory-boost

memory-boost

Memory for coding agents that reviews itself.

[CI] [License: MIT] [Python 3.11+]

Try it · What it answers · Connect your agent · How it works · vs. alternatives

Your agents write everything down. Nobody reads it back.

Every memory tool for coding agents helps them remember more. None of them mention that half of what they remember expired in July. memory-boost keeps a plain-markdown memory your agents read and write over MCP, and then does the unglamorous part: it audits that memory and your own sessions, and tells you what drifted. Politely, with numbers.

Think of it as a code review for what your agents remember and a fitness tracker for how you work with them. Mostly the fitness tracker: mine told me my most-failed command was cd. Sixty-nine times.

What it answers

Using an AI agent? Point Claude Code, Codex, Cursor or Gemini CLI at this repo: AGENTS.md tells it how to install memory-boost, seed a wiki and run the reports for you.

Try it in 30 seconds

No install needed; the wiki lands in ~/.local/share/memory-boost:

uvx memory-boost==0.2.1 init --example   # sample wiki with five kinds of rot + four synthetic transcriptsuvx memory-boost==0.2.1 drift            # what in that memory is stale, orphaned or never written downuvx memory-boost==0.2.1 mine             # now on your real ~/.claude/projects: aggregates only

To keep it in Claude Code, install the plugin. It brings the MCP server, the session-start brief and the skills in one step; it only needs uv on your PATH.

/plugin marketplace add arnaugonzalez/memory-boost/plugin install memory-boost@memory-boost

Other agents, or the CLI on its own: uv tool install memory-boost==0.2.1 (or pipx install memory-boost==0.2.1). Or let your agent do it: see Install from your agent.

Install from your agent

Paste this into Claude Code, Codex, Cline or Cursor:

Install memory-boost 0.2.1 by following https://github.com/arnaugonzalez/memory-boost/blob/main/llms-install.md : install the CLI with uv tool install memory-boost==0.2.1, run memory-boost init, register memory-boost serve as a stdio MCP server named memory, check that it answers with nine memory_* tools, then run memory-boost drift and show me the output. Ask me before writing outside the memory home (~/.local/share/memory-boost) or editing any global agent config.

The agent-oriented steps (stdio and serve --http, per harness, with checks) are in llms-install.md.

Example: back on a project after two months

You come back to acme-api and the agent still trusts everything it wrote down. drift says which of it no longer holds: a Postgres pin whose review date passed in June, a superseded decision that points nowhere, projects that went silent. page shows why the pin was made, so you can decide in a minute. lessons brings over what pixel-notes learned about retrying jobs, because both projects run background jobs. Bundled example data (memory-boost init --example), recorded with vhs from docs/demo.tape.

On a real machine

Five weeks of the author's own daily use: 29 decisions, 24 projects, 657 Claude Code transcripts. Output as produced; only project and file names are replaced.

$ memory-boost drift# Drift report — 2026-09-2429 decisions · 24 projects · 204 log entries · 33 open checkpoints · **20 findings**
## decision-expired (9)- deploy-via-bastion — review_by 2026-07-08 (78 days ago); revisit or mark historic- bot-guardrails — review_by 2026-07-08 (78 days ago); revisit or mark historic  …## project-no-decisions (2)- thesis-solver — 20 sessions logged, 0 decisions tagged 'thesis-solver': the learning lives only in transcripts## project-silent (5)- mood-api — last session 2026-04-16 (161 days ago); still listed as live  …
$ memory-boost mine# Session digest — 2026-09-24sessions      80 with ≥20 turnstools         Bash 5956 · Edit 860 · Read 692 · Write 493 · Agent 203 · ToolSearch 138bash          5956 calls, 159 failed (2.7%); most failed: cd (69), sleep (12), ssh (9), pkill (8), ls (7)bash/edit     4.4x; extreme: worktree-a 107.5x, sandbox 70.0x, sandbox 54.0xreads         692 calls, 131 re-reads of a file already read (19%); top: …/apply_email.py ×18discipline    62 long sessions (≥60 turns): 47 never checkpointed, 29 never savedsubagents     203 spawned in 44 sessions; max in one session: 20

Every line is a habit to change, a decision to revisit, or a rule worth writing down. None of it needed a new tool: the data was already on disk. It needed a reader.

What it is

1. A wiki you own. projects/*.md, decisions/*.md, concepts/*.md, log.md. Plain markdown; edit by hand, keep it in git. SQLite FTS5 index as a disposable cache. Nine MCP tools for any harness: memory_brief, memory_recall, memory_page, memory_save, memory_checkpoint, memory_resume, memory_drift, memory_lesson, memory_shelf. memory_shelf (CLI: memory-boost shelf "idea") appends a dated one-line idea to concepts/idea-shelf.md without the agent reading the page back.
2. Freshness as data. Decisions carry status: and review_by:. Search results are tagged current / unclassified / review / historic / superseded, current first.
3. Session checkpoints that survive /compact, with one line per other active session on the same project so parallel agents don't collide.
4. drift reads the wiki and reports what expired, what went silent, what was never written down and what was left half-done. The current project's top findings go into every session brief.
5. mine reads your local Claude Code transcripts and reports tool usage, failed commands, re-reads, subagent fan-out and checkpoint discipline. Aggregates only: no prompt, command or file content is stored or printed, so the digest is safe to paste anywhere.
6. lessons: concepts/ pages with applies_when: [tags]. A project page that lists matching tags: gets those lessons in its brief, so what worked in one repo reaches the next. The /retro skill writes them at session end.

Connect your harness

MCP server (all harnesses): command memory-boost, args ["serve"].

To share one server between several sessions, run it over streamable HTTP (stateless, JSON responses) and point each harness at http://127.0.0.1:8765/mcp:

memory-boost serve --http                      # --host 127.0.0.1 --port 8765 are the defaultsclaude mcp add --transport http memory -s user http://127.0.0.1:8765/mcp

It binds to loopback by default and has no authentication: keep it there.

Manual Claude Code hooks (~/.claude/settings.json), not needed with the plugin:

{ "hooks": {  "SessionStart": [ { "hooks": [ { "type": "command", "command": "memory-boost hook session-start" } ] } ],  "Stop":         [ { "hooks": [ { "type": "command", "command": "memory-boost hook stop", "timeout": 10 } ] } ],  "PreCompact":   [ { "hooks": [ { "type": "command", "command": "memory-boost hook pre-compact", "timeout": 20 } ] } ]} }

On startup/resume it injects ~1k tokens: catalog, the project's context and last sessions, its drift, open pending items, lessons from other projects, your open checkpoint. On compact only this session's checkpoint. It never blocks session start: on any error it prints nothing.

stop and pre-compact write an automatic checkpoint from the transcript, with no LLM call: your prompts (minus <private>…</private>), files edited and git commits, under the checkpoint's auto key. What the agent wrote (task, done, next) is never overwritten. stop writes only when something changed and 10 minutes or 4 prompts passed; pre-compact always writes, so the post-compact re-injection has something to show even if the agent never called memory_checkpoint.

Skills: /memory-boost (router: when to run what), /save-session, /restore-session, /retro. The plugin installs them as /memory-boost:retro and so on; without it, copy skills/* into ~/.claude/skills/. The first session after installing the plugin takes a few extra seconds while uv fetches memory-boost==0.2.1 and its dependencies from PyPI (the only network access the plugin makes); later sessions start in about 0.2 s.

CLI contract for agents: data on stdout, errors on stderr with a fix: line, --json on every reading command (with a version field), exit 0 = trust stdout, 1 = not found, 2 = bad input. Reports end with a next: line naming the exact command to run.

How it works

_{Source: docs/architecture.mmd}

The project is resolved from the working directory name (~/work/acme-api → projects/acme-api.md); map other names in aliases.json. Under 1,900 lines of Python, one dependency (mcp), no embeddings unless you opt into semantic recall, no daemon unless you run serve --http.

Session events v2

Every memory_save (or memory-boost save) writes one event to $MEMORY_BOOST_HOME/events/<agent>/<YYYY-MM-DD>.jsonl and one entry to wiki/log.md. The v2 fields are all optional: a v1 call (project, action, result, detail, pending as text) keeps working and writes the same log.md header. Each event now carries "schema": "memory-boost/session@2", a server date_local, a per-session seq and the assigned pending ids; v1 readers ignore the extra keys.

• outcome: done | progress | blocked | waiting | research | abandoned.
• pending: a list of {text, owner, severity, due?, waiting_on?}. The server assigns each item a stable id <project>#<n> (counter per project in events/pending-seq.json), shown in the brief under open pending.
• closes / drops: ids the session finished or no longer needs. Open items are what remains after folding them.
• session and metrics: optional blocks stored as given. The caller computes them, for example with a session summary tool; the save itself does not read transcripts.

{  "schema": "memory-boost/session@2",  "project": "acme-api",  "action": "Moved the job queue to Postgres and added retries",  "result": "ok",  "outcome": "progress",  "session": {"id": "3f9c2a1e-…", "harness": "claude-code", "cwd": "/home/me/work/acme-api",              "segment_start": "2026-10-28T10:00:03Z", "segment_end": "2026-10-28T14:45:12Z"},  "metrics": {"source": "ledger session-summary@1", "minutes": {"acme-api": {"attention": 182, "agent": 160, "gross": 230}},              "commits": [{"repo": "acme-api", "sha": "f2c1114", "pushed": false}],              "tests": {"runs": 6, "passed": 1180, "failed": 0}, "cost_usd": 14.2},  "pending": [{"text": "Run the batch import", "owner": "user", "severity": 3},              {"text": "Push main to origin", "owner": "user", "severity": 2}],  "closes": ["acme-api#12", "acme-api#14"]}

date_local is set by the server from MEMORY_BOOST_TZ (an IANA zone such as Europe/Madrid; UTC by default). Events written before v2 have no pending ids, so memory-boost pending [--project] (and closes) cannot see them.

Semantic recall (optional)

Off by default: memory_recall is full-text only and the base install stays one dependency. To also match paraphrases and other languages, run a local embedding model:

pip install "memory-boost[semantic]==0.2.1"   # adds sqlite-vecollama pull embeddinggemmaexport MEMORY_BOOST_SEMANTIC=1memory-boost index                       # embeds the wiki once; later runs embed only changes

The first index of a large wiki on CPU can take minutes: as a rough guide, about 15 minutes for some 850 documents on a laptop without a GPU (one measurement, not a benchmark). memory-boost index and recall show embedding N/M chunks… on stderr while it runs; later runs re-embed only what changed.

Default mode when on is hybrid (full-text and vectors fused by weighted reciprocal rank); override with MEMORY_BOOST_RECALL_MODE=fts|semantic|hybrid or memory-boost recall --mode. If Ollama is down, recall falls back to full-text and logs a warning. On a private 64-query set (39 literal, 25 paraphrase) with a warm index on CPU: paraphrase nDCG@10 0.46 to 0.76, literal 0.80 to 0.85, p50 about 110 ms per query. The query set is not published.

Configuration

Status values: current, review, historic, superseded; accepted/active read as current and proposed as review.

Experimental (shipped, little real use yet): memory_shelf, memory-boost friction and the lesson fields cue/evidence/confidence. Their output format may still change.

vs. alternatives

Want better search, more integrations and a community? Use basic-memory. Want everything captured automatically? claude-mem. Use memory-boost if you want to learn from what your agents already wrote down.

Limits / non-goals

• Full-text search by default; embeddings only with the optional [semantic] extra and a local Ollama.
• mine understands Claude Code transcripts today. Codex is on the roadmap.
• Reports are as good as the wiki: an empty wiki yields an empty drift.
• Single user, local files. Linux and macOS; Windows untested.
• The wiki is read by your agents as instructions. Only use a wiki you trust; see SECURITY.md.

License

MIT