gitmem

Persistent, reviewable memory for your AI agents — in a git repo you can read, diff, and blame.

Your coding agent forgets everything between sessions. gitmem gives it an append-only event log of facts, decisions, and corrections, stored as plain JSONL in git, with deterministic projections: a token-budgeted brief to inject into context, a current-facts view, and a conflict queue that surfaces contradictions instead of silently overwriting them.

No vector store. No LLM calls. No server. A memory system you can git log.

<p align="center"><img src="assets/demo.svg" alt="gitmem demo" width="780"></p>

Installation

Install from npm:

npm install -g @josephy02/gitmem

Or, if you are developing or setting up the Claude Code plugin, clone the repository and install it locally:

git clone https://github.com/josephy02/gitmem.git
cd gitmem
npm install   # builds automatically
npm link      # puts `gitmem` on your PATH

Verify the install:

gitmem --help

60-second quickstart

gitmem init --root ./memory

gitmem --root ./memory append --scope team/core --kind decision \
  --body "Mobile still depends on the old auth module; do not refactor." \
  --author human:joseph

gitmem --root ./memory append --scope team/core \
  --body "The staging DB is reset every Sunday 03:00 UTC." \
  --author agent:builder-3

gitmem --root ./memory brief         # the context bootstrap, capped at 1,500 tokens
gitmem --root ./memory facts --json  # current-value view, NDJSON
gitmem --root ./memory conflicts     # contradictions, surfaced never auto-resolved
gitmem --root ./memory commit        # git commit of the log, on your cadence

Or explore the bundled demo — 45 realistic events with corrections, a retraction, a promotion, and a live conflict:

gitmem --root /tmp/demo init
gitmem --root /tmp/demo append --json --force - < demo/events.ndjson
gitmem --root /tmp/demo brief

How it works

flowchart LR
    subgraph writers[" "]
        CLI[CLI / library]
        MCP[MCP client<br/>Claude Code etc.]
    end
    CLI -->|append| LOG
    MCP -->|memory_append| LOG
    LOG[("log/YYYY/MM/DD.jsonl<br/>append-only, in git")]
    LOG -->|pure function| PROJ[projections]
    PROJ --> BRIEF["brief.md<br/>≤1500 tokens"]
    PROJ --> FACTS["facts.json<br/>live/superseded/contested"]
    PROJ --> CONF["conflicts.json<br/>never auto-resolved"]
    LOG -.->|every read| CHOKE{{"readEvents()<br/>capability choke point"}}
    CHOKE --> BRIEF & FACTS & CONF
    GIT[git history] -->|"gitmem stale"| FACTS

1. The log is the only source of truth. One event per line in log/YYYY/MM/DD.jsonl. Nothing is ever mutated or deleted — corrections and retractions are new events that supersede old ones, so provenance is always reconstructible (gitmem trace <id>).
2. Projections are pure functions of the log. facts.json (current values with live/superseded/retracted/expired/contested status), brief.md (the always-injected core, hard-capped at 1,500 tokens, decisions first), conflicts.json, stats.json. gitmem rebuild is byte-identical to an incremental build — that's a test.
3. Conflicts are surfaced, never auto-resolved. Deterministic heuristics (divergent corrections, negation pairs, same-subject divergence) flag contradictions; both sides are returned together as contested. Resolution is a human act: write a correction that supersedes the losers.
4. Scope is enforced at one choke point. Every read path — search, point-get, brief, trace — goes through a single capability-checked function. Segment-aware: team/core grants team/core/auth but never team/core-secrets. Promotions change a fact's effective scope, and access control follows the effective scope, so narrowing actually narrows.
5. Git-native for real. gitmem init installs a union merge driver: two branches appending to the same day file merge automatically — union of lines, sorted by ULID, always correct because events are immutable. gitmem verify catches duplicate ids from bad merges.

Event format

The format is the product. One JSON object per line, schema in schema/memevent.schema.json — any language can write events without this library:

{"id":"01K2X9...","ts":"2026-08-15T14:03:11.000Z","scope":"team/core","author":{"kind":"human","id":"joseph"},"kind":"decision","body":"Mobile still depends on the old auth module; do not refactor.","derived_from":[],"supersedes":[],"confidence":1}

Five event kinds: observation, decision, correction, retraction, promotion (scope changes are events too — sharing has provenance).

Library

import { GitMem } from "@josephy02/gitmem";

const log = GitMem.open("./memory");
const cap = { principal: "agent:builder-3", scopes: ["team/core"], mode: "read" as const };

log.append({ scope: "team/core", kind: "observation", body: "...", author: { kind: "agent", id: "builder-3" } });
log.brief(cap);        // markdown string, reprojects lazily if the log advanced
log.facts(cap, { status: "live" });
log.conflicts(cap);
log.trace(cap, id);    // full derivation ancestry

Design commitments

• No LLM in the write path. Writes are cheap, lossless, synchronous.
• No write-time dedup. Contradictions look like near-duplicates; a write-time gate would reject exactly the events the conflict detector needs to see. Everything is admitted; resolution happens at projection time.
• brief.override.md — a human-authored file that always wins the top of the brief.
• Human-first storage. git diff a memory change. git blame a fact. Review an agent's memory in a PR.

Claude Code plugin

The fastest way to give Claude Code persistent memory. This repo is a plugin marketplace:

/plugin marketplace add josephy02/gitmem
/plugin install gitmem@gitmem

(Requires the gitmem CLI: npm install -g @josephy02/gitmem.)

What you get:

• Memory brief at session start — a SessionStart hook injects gitmem brief into context, so every session begins knowing your project's decisions and facts. No gitmem root in the project? The hook is a silent no-op.
• Memory tools over MCP — Claude can append observations, decisions, and corrections as it works. The root is auto-discovered ($GITMEM_ROOT, ./.gitmem, ./memory, ./.memory) and auto-initialized on first use.
• /remember <fact> — save a durable fact or decision, with correction semantics when it contradicts an existing memory. /remember with no arguments harvests the current conversation.
• /memory-review — walk the conflict queue and stale anchors, and resolve them through the log.

MCP server

Give any MCP client (Claude Code, Claude Desktop, anything speaking MCP) persistent memory in one line:

{
  "mcpServers": {
    "gitmem": { "command": "gitmem", "args": ["--root", "/path/to/memory", "serve"] }
  }
}

Exposes five tools over stdio: memory_append, memory_brief, memory_facts, memory_conflicts, memory_trace. Appends are attributed to agent:mcp by default (--author to change); reads go through the same capability choke point as everything else.

Git-anchored staleness

A fact can anchor itself to code via meta.source_uri (e.g. "src/auth.ts#validateToken"). Because the log lives in git next to the code, staleness detection is just a git log:

gitmem stale            # lists live facts whose anchored file changed since the fact was written

[stale?] validateToken always returns true in dev mode
  anchor: src/auth.ts#validateToken
  changed by:
    e1faa27 flip validateToken default

No embeddings, no LLM, no index to maintain — the same property that makes memory reviewable makes it self-invalidating.

Development

npm install
npm run build
npm test        # 16 tests incl. property-based scope isolation and a real git-branch merge

Performance: full projection of a 10k-event log runs in ~50ms.

License

MIT