memgit

[memgit logo]

memgit — git for AI memory

Your AI assistants forget everything when the session ends. memgit fixes that.

Version-controlled, cross-AI context that persists, diffs, rolls back, and syncs like code. Switch from Claude to Cursor to ChatGPT mid-project — your context is already there.

[PyPI] [License: MIT] [Tests]

Why not claude.md? Why not mem-search?

You've probably already tried both. Here's why they hit a ceiling:

Proof — context costs you can measure

Run this on your own store to see the actual numbers (measured where possible; estimates labeled):

$ memgit stats
  Total memories:   108   (41 feedback · 23 user · 19 project · 12 reference · 8 convention · 5 lesson)  Priority:          3 critical · 67 medium · 38 low
  Context footprint  (measured where possible; estimates labeled)
  Surface                                             Tokens  Full store (every memory as context)                12,840  Resume digest  (measured render)                       540  Recall block  (estimate: top-3 rules ≈ chars/4)        ~60
  per-session injected ≈ 600 tokens (estimate) vs 12,840 tokens if the full store were loaded

Why such a big difference? claude.md loads all context every session. memgit injects a bounded resume digest plus BM25-matched recall — only what is relevant to this session, not everything you've ever recorded. The digest is measured by actually rendering it, and the store total is the real corpus size; nothing here is a simulated benchmark.

The git analogy is literal

memgit's data model maps exactly to git:

This is not metaphorical — memgit uses a content-addressed object store (SHA-256 blobs) identical to git's architecture. Every memory has a stable SHA. Identical content has identical SHAs. Old state is always recoverable.

The store IS a git repo

Every memory is a readable .toon file under memories/. Push your entire memory set to GitHub with standard git:

memgit git init --remote [email protected]:yourteam/ai-memory.gitmemgit git push

Teammates pull and start with your AI's learned rules from session 1:

git clone [email protected]:yourteam/ai-memory.git ~/.claude/memgit-storememgit setup all

You can grep, git blame, and git diff your memories just like code:

grep -rl "database" ~/.claude/memgit-store/memories/git log --follow memories/no-db-mock.toongit diff HEAD~7 memories/

Install

Mac / Linux:

pip install memgit

Mac (Homebrew):

brew tap code4161/tap && brew install memgit

Windows:

choco install memgit# orpip install memgit

(The Chocolatey package is live on community.chocolatey.org; newly pushed versions can take a few days to clear moderation — pip install memgit always has the latest.)

Any AI tool config (no Python needed — npx auto-installs on first run):

{ "mcpServers": { "memgit": { "command": "npx", "args": ["-y", "memgit-mcp"] } } }

Quickstart (3 minutes)

# 1. Install and initializepip install memgitmemgit init               # auto-detects the best location, finds your existing                          # Claude Code memories, and offers to import them
# 2. Register with your AI tools (interactive picker)memgit setup
# 3. See your token savingsmemgit stats

init walks you through it — no paths to hunt down. (Importing later is one command with no arguments: memgit sync auto-finds ~/.claude/projects/*/memory.)

Restart your AI tool — it now searches your memory store at the start of every session.

Adopting memgit mid-project

Memory tools have a cold-start problem: install one halfway through a project and it knows nothing — there's no initial point, and context only trickles in from future sessions. memgit solves this with a one-time seeding pass:

cd your-projectmemgit onboard          # mines the repo, prints the bootstrap brief

onboard first extracts a repo digest deterministically — git history (recent commit subjects, hot files/directories by churn, authors, branch, tags), detected stack from manifests, and the docs worth reading — using bounded, read-only probes that stay near-instant even on huge repositories. The brief then tells your AI agent exactly what to do with it: read only the listed files (no tree crawling), extract 10–20 durable facts (purpose, architecture, conventions, current state, gotchas), save each as a typed memory, and checkpoint the seed set. Paste it into a session — or don't: if the AI searches memory in a project that has none, the MCP server itself replies with the bootstrap instructions instead of a bare "no results."

Memories are project-scoped, filter-by-default (v0.7.0): each carries the workspace it belongs to, and searches, recall injections, and the resume digest (recent memories, checkpoints, depth hints) are filtered to the current project's family plus explicitly-global memories — another project's content never leaks in. Widen deliberately with memgit search --all-projects / all_projects: true (every hit then carries its project label), or hard-filter one project with --project. A memory with no project is explicitly global (applies everywhere): save one with memgit add --global or project: "". A save whose project cannot be determined is never silently global — it's quarantined under _unknown (visible in list as [?project], flagged by lint, surfaced nowhere) until you relabel it with memgit doctor --relabel.

Resume where you left off

Ask an AI "can we proceed on the pending tasks?" in a fresh session and it will guess from whatever file happens to be open. memgit resume replaces the guess with the record:

memgit resume            # last checkpoints, work in flight, recent + critical memoriesmemgit resume --plain    # plain text, for piping into an AI contextmemgit resume --json     # for tooling

Wire it into Claude Code so memory becomes automatic — no tool call, no judgment required:

memgit setup hooks       # installs all six hooks (~/.claude/settings.json)

Why hooks and not just good tool descriptions? We measured it: across 166 real sessions, hook-injected context was delivered in 100% of them while voluntary memory-tool calls happened in 6%. What a hook enforces happens.

The resume digest is deliberately bounded (~350 tokens measured on a 500-memory store): rules are clipped, the critical list is capped, and full text is one get_memory call away.

Scale to 10,000+ sessions

After months of use, your checkpoint history grows. Squash compresses it, gc reclaims the disk:

memgit squash --keep-last 100    # keep last 100 checkpoints, squash everything oldermemgit squash --older-than 30    # squash everything older than 30 daysmemgit squash --dry-run          # preview first
memgit gc                        # delete unreachable objects, trim reflogsmemgit gc --dry-run              # previewmemgit gc --squash-keep 200      # compact history, then sweep

The current memory state is always preserved — and squash is lossless-in-substance: every collapsed checkpoint leaves a one-line record (time, author, diff, message) in an append-only archive under .memgit/logs/archive/ that gc never touches. Benchmark on a 2,000-checkpoint store: 94% smaller (39.5 MB → 2.2 MB), fsck clean. History operations stay O(1) as the chain grows (SHA resolution and checkpoint counting measured at ~0.08 ms at 2,000 checkpoints).

Multiple agents, one memory

All writes go through a git-style store lock (0.08 ms overhead), so concurrent agents can't corrupt the store or lose each other's updates. Two patterns:

Shared thread — agents write concurrently; if one commits while another has work staged, the second commit auto-merges (three-way, against the recorded base) instead of clobbering. Set MEMGIT_AUTHOR=agent-name so each checkpoint says who did it.

Thread per agent — isolate, then integrate:

memgit thread create agent-1     # branch off for each agent# ... agents work on their own threads ...memgit merge agent-1             # three-way merge back (common-ancestor based)

Conflicts (same memory changed on both sides) resolve to the newest version; an edit always beats a delete. Both histories are preserved.

What the AI sees

Once registered via MCP, every AI tool gets 6 tools:

The tool descriptions teach the AI judgment — "does this request depend on state you don't have in context?" — rather than keyword triggers. Measured cost of the whole tool surface: ~1,150 tokens once per session; a resume_session reply is ~335.

A save that arrives damaged is refused, not stored (v0.13.0). A model sometimes closes a value with a tag named after the field (</rule>), and the host then delivers the next parameters inside that value: why, when, tags and type_code never arrive, and the memory is stored with defaults. On one machine that was 748 of 3,025 saves. save_memory now refuses any value carrying tool-call markup, names the field and the parameters it swallowed, and the model retries in the same turn. Markup quoted on purpose inside backticks is allowed. Every successful save echoes the length of each field it stored, which fields arrived empty, and which values came from a default. Damaged memories already in a store are never served back into context: the digest, the recall hook and read results show each field cut at the damage. memgit doctor counts them, and memgit doctor --repair-markup splits them back into their fields.

Core operating guide (v0.5.0)

A project's hardest onboarding problem isn't what it does — it's how to work in it: which skill to invoke, which command to run, which tool to reach for. That lives in a CLAUDE.md or a skills folder the AI host may or may not be configured to read. memgit carries it for you.

memgit core seed distills a compact operating guide from the project's existing skills + rule files. memgit core sync writes it into every AI host's own rules surface as a dedicated, memgit-owned file — .claude/rules/memgit.md, .cursor/rules/memgit.mdc, .windsurf/rules/memgit.md, .clinerules/, .roo/rules/, .continue/rules/ — and marker-delimited blocks in the shared GEMINI.md (Gemini CLI auto-loads only that file) and Codex's AGENTS.md. It's additive only — memgit never touches your own config or content — and injected at session start, so any tool knows how to work in the project even when its native setup is missing.

And it learns: a sidecar usage ledger tracks which memories actually get recalled, and the most-used ones are auto-promoted as pointers into the guide over time (budget-capped, decaying, and always subordinate to the repo's own rules — it never restates or overrides them). Drifted? memgit core heal rebuilds it.

Since 0.8.0 it also starts itself: a project's first guide is created automatically once it holds a handful of memories, so the loop no longer waits on someone remembering to run core seed. The guide leads with what the project actually holds — "this project has N saved memories covering topics" — because a stated count of real prior work is evidence a model can act on, where an abstract instruction to check memory is something it can weigh against its own confidence and skip.

memgit Pro — one memory across every machine

The local engine, the store, the MCP server and every command below are MIT and never gated. Pro is the hosted layer for people who work on more than one machine or with a team: end-to-end-encrypted sync where the server stores ciphertext only, unlocked by one licence key.

memgit pro activate <key>     # validate the key with Polar and store it (mode 0600)memgit pro status             # Free / Pro, last verification, expirymemgit pro deactivate         # remove the key from this machine

• Buy a key at memgit.dev/#pricing — $12/month or $99/year, billed by Polar (merchant of record, taxes handled). The key arrives by email and in your Polar purchases page.
• The only bytes that leave your machine are the key and memgit's public organisation id, sent to Polar's validation endpoint. No memory content is ever sent.
• Headless hosts: set MEMGIT_LICENSE_KEY=<key> in the MCP server's environment; it is validated with the same cache and never written to disk.
• Fail-open: a key that verified in the last 14 days keeps working offline; a rejected key drops to Free with a clear message. A lapsed plan never locks your data — reads, pulls and exports keep working.
• If this store is logged in to memgit cloud, pro activate also upgrades that account.

Sync between machines and with a team

memgit cloud signup --email [email protected]   # 14-day trial; keys are made on this machinememgit cloud team create myteammemgit cloud link myteam/memory               # creates the cloud repo on first linkmemgit cloud auto on                          # every save syncs in the background from now on

With autosync on, each save (including every save_memory your AI makes) starts one background pull-then-push and returns immediately, and each session start pulls what other machines saved. memgit cloud auto status shows the last run and why it would not run.

On a machine that runs an agent but no interactive sessions (a Mac mini, a VM), log in and link once, then keep it current under systemd or launchd:

memgit cloud watch --interval 60                       # pull and push forever, one JSON line per runmemgit cloud watch --export ~/agent/.hermes.md         # also write the digest for agents that read context files

Everything leaves the machine end-to-end encrypted; the server stores ciphertext only.

Backups that actually happen

memgit's premise is that the AI is the operator — but backup used to require a human to remember memgit git init --remote <url> and keep pushing. On this project's own store that meant 1,734 memories on one disk with no copy anywhere, five weeks in. A maintenance task that needs a human command is a task that will not happen.

Since 0.9.0 it runs itself, at the end of a session, with the safety boundary drawn at network egress rather than effort:

• Local destinations are automatic — a cloud-synced folder you already have (iCloud, Dropbox, Google Drive, OneDrive) or an external volume. memgit copies files; it opens no connection and signs up for no service. Your existing sync client does the rest.
• A git remote is pushed to only if you already configured one. memgit never invents a remote, never creates a repository, and never sends memories to a host you did not choose — memories can contain credentials, and convenience is not a reason to publish them somewhere you never picked.

The backup is a single memgit-store.tar.gz, not a directory tree: a 203 MB store is 10,295 small object files, and giving a sync client 10k files to reconcile every time is how you get a sync client that never finishes. It is staged and renamed atomically, keeping the old copy until the new one lands — an interrupted backup must never leave a corrupt file where a good one used to be.

memgit backup status     # where the last copy went, and what else is available

Ranking you can prove

Retrieval quality used to be adjusted on intuition. memgit eval replaces that with a measurement, using two frozen sets mined from the store itself:

• recall — real prompts and the memories memgit actually surfaced for them. Measures stability: did a change break what used to work?
• synthetic — each memory queried by its own why, expecting itself back, with the slug's own words stripped out of the query. Measures correctness, independently of any past ranking. (The recall set alone is circular — its answers came from the ranking under test.)

Both report hit@1, recall@3/5/10 and MRR against a pinned baseline. There is deliberately no single blended score, and no fabricated "tokens saved" number.

It earns its keep immediately. Two changes built for 0.8.0 looked obviously right and were measured wrong: a recency multiplier (cut — real-prompt hit@1 −0.020, MRR −0.018) and destructive stemming, which fixed its motivating query while costing hit@1 −0.038 overall (rebuilt as an additive field, then +0.020/+0.041/+0.019). A BM25 normalisation bug introduced in the same release was caught the same way.

Depth advertisement, trackers & supersession (v0.6.0)

Measured across 289 real sessions: injected recall reached ~59% of them, but only 6.8% ever ran an active search — the injected top-3 reads as "memory consulted", so the model never learns there's a queryable store behind it. 0.6.0 makes the passive layer advertise what the active layer knows:

• Memory index — the resume digest ends with tag→count pairs (8a8f4ec (6) · instagram (5)) and the exact call to go deeper. Counts are truthful: superseded memories are excluded, and every advertised topic is guaranteed to return search results.
• "+N more" recall hints — when the per-prompt recall block has more on-topic memories behind it, it says so, with the one call to get them.
• Context-triggered recall — a PostToolUse hook: reading a file whose path matches a memory tag surfaces memgit: 6 memories tagged 'x' relate to this path. Reads only a commit-time tagmap cache (never the store), capped 3/session.
• Trackers (tr) — one memory per in-flight entity (<entity>-status), updated by re-saving the same slug. They render as a status board at the top of every session: memgit is the authority for entity status; files may lag.
• Supersession — a correction names what it replaces (supersedes=[old-slug]) instead of a "CORRECTED:" prefix. Superseded memories vanish from search/recall/resume (history preserved; list still shows them marked ⊘), so injected context is never stale.

Commands

# Core (git-like)memgit init                       # initialize store (auto-detects best path)memgit onboard                    # bootstrap brief for an existing codebasememgit add <slug> <rule>          # stage a memory (--body detail, --project scope, --global everywhere, --supersedes old-slug)memgit commit -m "message"        # checkpoint current statememgit log                        # historymemgit diff [sha1] [sha2]         # what changedmemgit show <slug>                # display a memorymemgit remove <slug>              # remove from active index (history preserved)memgit status                     # staged changesmemgit search <query>             # BM25 search, scoped to this project + global (--all-projects to widen)memgit rollback <ref>             # restore state to a checkpoint (HEAD~N or SHA)memgit resume                     # where we left off — session-start digestmemgit merge <thread>             # three-way merge a thread into the current onememgit remove <slug>              # (aliases: delete, rm, del) — mistypes get a "did you mean?"
# Core operating guide — per-project, always-on, cross-hostmemgit core seed                  # draft a guide from this project's skills + rule filesmemgit core sync                  # deliver it into each AI host's own rules file (additive)memgit core show / edit           # view / curate the guidememgit core heal                  # self-repair a guide that has drifted                                  # (a project's FIRST guide is created automatically                                  #  once it holds 5+ memories — no command needed)
# Durability — automatic, no human command requiredmemgit backup status              # last copy, staleness, available destinationsmemgit backup now                 # force one immediatelymemgit backup set <path>          # pin a destinationmemgit backup off / on            # control the automatic path
# Retrieval evaluation — prove a ranking change helpedmemgit eval mine                  # freeze a regression set from real recall eventsmemgit eval mine --synthetic      # non-circular set: query each memory by its own `why`memgit eval run --set recall      # hit@1 / recall@3,5,10 / MRR vs the pinned baselinememgit eval run --baseline        # pin the current result as the comparison pointmemgit eval run --misses 10       # inspect the cases where nothing relevant surfaced
# Scale & proofmemgit squash                     # compress old history (archives what it collapses)memgit gc                         # reclaim disk: sweep unreachable objects + stale session cachesmemgit stats                      # measured context costs + disk usagememgit doctor                     # hygiene + scope losses: split labels, quarantined, stale caches, orphaned usagememgit doctor --audit             # every split and stranded label, the save landing rate, throughputmemgit doctor --audit --json      # the same report as JSONmemgit doctor --relabel map.json  # bulk re-project memories ({"slug": "Label" | ""}); one checkpointmemgit doctor --repair-markup     # dry run: memories whose fields swallowed other parameters on savememgit doctor --repair-markup --yes [--project LABEL]   # split them back into their fields; one checkpointmemgit lint                       # validate all memories (flags unknown provenance)memgit fsck                       # verify store integrity
# Import / exportmemgit sync                       # sync from Claude Code files + commit (auto-finds them)memgit import claude-code [path]  # path optional — defaults to ~/.claude/projects/*/memorymemgit import file <path>memgit export <slug>
# Git sync (team features)memgit git init [--remote URL]memgit git push [remote] [branch]memgit git pull [remote] [branch]memgit git exportmemgit git status
# AI tool registrationmemgit setup                      # interactive step-by-step pickermemgit setup all                  # auto-register every detected toolmemgit setup claude-codememgit setup cursormemgit setup windsurfmemgit setup clinememgit setup continuememgit setup gemini-climemgit setup hooks                # Claude Code hooks: resume at start, per-prompt recall,                                  # action gate on Bash and file writes, compliance and                                  # capture guard + auto-sync at stop (--no-recall / --no-guard / --no-gate)memgit doctor --rules             # critical rules next to same-topic lines in CLAUDE.md, AGENTS.md, ...
# Pro (licence key)memgit pro activate <key>         # Polar-issued key; unlocks hosted E2E syncmemgit pro status / deactivate
# Servermemgit serve                      # MCP stdio (Claude Code, Cursor, Windsurf, Cline)memgit serve --http               # HTTP REST (ChatGPT Custom Actions, Gemini)
# Visualizationmemgit graph                      # D3.js interactive relationship mapmemgit thread list / switch / create

AI tool support

What memgit reads from your environment

memgit runs inside your agent, so the honest answer to "what does this process read from my environment?" should be checkable rather than a promise. The table below is generated from memgit/env.py, and tests/test_env_inventory.py walks the package on every test run and fails if the code reads a name this list does not carry — or carries a name nothing reads. The list cannot drift away from the code.

memgit reads these and nothing else. It writes none of them. None of them turn network access on: the cloud endpoints are overrides for a sync you have already opted into with memgit cloud login, and with no licence and no cloud login, memgit talks to nothing.

How the project label is decided

Memories are project-scoped, so a wrong label files a memory where you will not find it again. The label comes from the first of these that yields one, and there is exactly one detection path shared by the MCP server, the CLI and the hooks — so a label derived at save time and one derived at recall time cannot disagree:

1. MEMGIT_PROJECT — a forced label, taken verbatim
2. the cwd a host reports in its hook payload (the real workspace, even when the process working directory is elsewhere)
3. CLAUDE_PROJECT_DIR
4. the process working directory

The label keeps [A-Za-z0-9-] and turns everything else into a dash, exactly as Claude Code names its projects/ directories — including the _ character, which becomes a dash. Labels written before v0.12.0 kept the _; they are folded onto the dash form when two labels are compared, so nothing written under the old form stops resolving, and nothing stored is rewritten.

A project whose memories have split across two labels raises no error anywhere: the save succeeds, fsck stays clean, and the only symptom is an answer that is missing things. memgit doctor reports splits, stranded labels and the save landing rate by default; memgit doctor --audit adds every row. A repair heals toward the label whose directory exists, never toward the one holding more memories.

When none of them yields a label — running from $HOME, for instance — a write is quarantined under _unknown rather than filed globally, and memgit doctor --relabel moves it once you say where it belongs. Folding a free-text label onto an existing project (log-report → Downloads-log-report) happens only when the input is a trailing segment of exactly one known label and that label already holds memories; anything ambiguous is left alone, and every fold is reported rather than done silently.

TOON format — compact, readable, diffable

Standard markdown memory file:

## Rule: Never mock the database in tests**Type:** feedback  **Priority:** medium  **Why:** We got burned last quarter — mocked tests passed but the prod migration failed.  **When to apply:** Any time writing tests that touch persistence layers.  **Tags:** testing, database

The same memory in TOON:

TOON1|fb|no-db-mock|2026-07-01T10:00Z#testing #databasePROJ:my-appRULE:Never mock the database in testsWHY:Mocked tests passed but prod migration failed last quarterWHEN:Any persistence testBODY:Full long-form detail lives here, losslessly (newlines escaped).\nSearch returns the compact RULE; get_memory returns everything.

Measured with a real tokenizer, TOON is ~5–10% leaner than equivalent markdown — a nice bonus, not the headline. The headline saving is retrieval: memgit loads the top-8 relevant memories per query instead of everything.

At 108 memories: 12,840 tokens (dump everything) → 640 tokens (memgit BM25 top-8)

For exact token counts in memgit stats, install the optional tokenizer: pip install "memgit[tokens]".

Architecture

~/.claude/memgit-store/  .memgit/    objects/     ← SHA-256 content-addressed blobs (gzip compressed)    refs/threads/main   ← HEAD checkpoint SHA    TOON_INDEX   ← active slug→sha mapping    config       ← author, default thread    logs/        ← ref change audit trail  memories/      ← flat .toon files (git-trackable, human-readable)  .git/          ← standard git repo (after `memgit git init`)

Contributing

git clone https://github.com/code4161/memgit.gitcd memgitpython -m venv .venv && source .venv/bin/activatepip install -e ".[dev]"pytest    # 431 tests, all passing, < 5 seconds

See CONTRIBUTING.md.

Roadmap

• Content-addressed object store (git-identical architecture)
• TOON format (compact line-oriented memory format)
• MCP server — Claude Code, Cursor, Windsurf, Cline, Continue.dev
• HTTP server — ChatGPT Custom Actions, Gemini function calling
• BM25 relevance search (load only what matters)
• memgit stats — measured token savings proof
• memgit squash — scale to 10k+ sessions
• memgit git push/pull — team sync via standard git
• Flat memories/ directory — grep/diff/blame your memories
• D3.js graph visualization of memory relationships
• memgit resume + SessionStart hook — sessions start with "where we left off"
• Guardrail hooks — per-prompt auto-recall + end-of-session capture guard (v0.4.0)
• Core operating guide — per-project, always-on, cross-host, self-improving (v0.5.0)
• memgit gc — space reclamation (mark-and-sweep, lossless squash archive)
• Multi-agent write safety — store lock, auto-merge commits, memgit merge
• PyPI + Homebrew (tap) + npm published (v0.1.5)
• Chocolatey — live on community.chocolatey.org (choco install memgit)
• Interactive setup wizard (memgit setup)
• Smart memgit init (auto-detects tool, no path needed)
• Lossless memories — full body alongside the compact rule (v0.3.0)
• Project-scoped memories + memgit onboard mid-project bootstrap (v0.3.0)
• VS Code extension (v0.1.5, Marketplace: code416-memgit.memgit)
• Codex + Antigravity support — MCP registration + AGENTS.md core guide (v0.8.0)
• memgit eval — measured retrieval quality, real + non-circular sets (v0.8.0)
• Usage-aware ranking — the recall ledger feeds relevance, not just the guide (v0.8.0)
• Automatic core-guide bootstrap — the self-improving loop starts itself (v0.8.0)
• Automatic off-machine backup — no human command required (v0.9.0)
• JetBrains plugin (Phase 3)
• Semantic search via embeddings — gated on memgit eval showing a real gain (Phase 4)
• Public benchmark numbers (LongMemEval, LoCoMo) (Phase 4)
• memgit.dev website (live)
• Memory compression / auto-summarization (Phase 5)
• Team access control + audit trail (Phase 5)
• Memory marketplace — share reusable context packs (Phase 6)

License

MIT — see LICENSE.