# mement

> A local memory layer for coding agents. Your agents forget everything; mement doesn't. It reads the sessions of Pi, Codex, and Claude Code, keeps the words you actually said, and lets you search them. No cloud, no server — memory stays on your machine.

## What it is

mement is a local-first conversation memory CLI for coding agents. It ingests session history from Pi, Codex, and Claude Code, stores it in a local SQLite database, and provides full-text search over past conversations (including Chinese, via FTS5 trigram tokenization).

- Local-first: everything lives in a single SQLite file on your machine
- Zero runtime dependencies: built on Node's built-in `node:sqlite` (requires Node >= 22.13)
- Incremental ingestion: `mement add` resumes at the last complete log boundary
- Search: BM25-ranked full-text search with context windows, filterable by harness and project directory

## Install

```bash
npm install -g @hidaviddong/mement
mement install
```

`mement install` drops a skill into each agent:
- `~/.pi/agent/skills/mement/SKILL.md`
- `~/.codex/skills/mement/SKILL.md`
- `~/.claude/skills/mement/SKILL.md`

## Usage

Remember (ingest conversations incrementally):

```bash
mement add                    # all sources
mement add --pi               # just one source
mement add --codex --claude-code
```

Recall (search your history):

```bash
mement search "sqlite schema"
mement search "deploy failed" --cwd "$(pwd)" --context 3
mement search "query" --harness codex --limit 10 --json
```

## Details

- Memory lives at `~/.mement/mement.db`; set `MEMENT_HOME` to move it
- Storage: SQLite (WAL) with tables for sessions, messages, a trigram FTS5 index, and per-file ingest cursors
- Ranking: short queries (< 3 chars) use position-first LIKE ranking; longer queries blend FTS5 bm25 with literal hit counts
- Release: pushing any `v*` tag runs the GitHub Actions Release workflow (tests, typecheck, build, npm publish via OIDC, GitHub Release)

## License

AGPL-3.0-only

## Links

- GitHub: https://github.com/hidaviddong/mement
- npm: https://www.npmjs.com/package/@hidaviddong/mement