@rememlyai/hooks - Rememly capture hooks

Capture AI coding sessions into your personal Rememly memory layer, so the next session (in any tool) starts already knowing what happened in the last one. This is the capture layer of Rememly's double-layer memory model (ADR-018): raw, verbatim, personal-only, and hash-idempotent. The curated layer (approval-gated extractions) is built separately by the extract_memories tool.

Two guarantees:

• Personal-only. Captured transcripts are written to your personal scope and never to a team scope — a raw, unreviewed transcript is never pushed at a teammate. The server rejects any non-personal capture.
• Idempotent. Every chunk is content-hashed, so re-uploading an overlapping transcript (the hooks fire repeatedly) saves nothing new.

Configure

Set these in your shell environment (the hooks read them at fire time):

| Variable | Required | Default | Meaning | | --- | --- | --- | --- | | REMEMLY_TOKEN | ✅ | - | A cm_* API token with memory:write | | REMEMLY_API_URL | | https://api.rememly.ai | Backend base URL | | REMEMLY_CAPTURE_EVERY | | 15 | Human turns between Stop uploads | | REMEMLY_CAPTURE_OFF | | - | Set to disable capture entirely | | REMEMLY_CAPTURE_DEBUG | | - | Set to log hook activity to stderr |

Claude Code

Full transcript capture. Install as a plugin (the .claude-plugin/ manifest wires the Stop, PreCompact, and SessionEnd hooks):

npm i -g @rememlyai/hooks
# then add this repo's plugin dir to Claude Code, or copy .claude-plugin/hooks.json
# into your ~/.claude/settings.json hooks block.

• Stop - uploads once every REMEMLY_CAPTURE_EVERY human turns.
• PreCompact — always uploads (the transcript is about to be summarised away; this is the one hard-guarantee firing).
• SessionEnd — uploads anything new before the session closes.

The hook never emits blocking output — it always exits 0, so it can never wedge Claude Code, and it never loops on stop_hook_active.

Cursor

Cursor doesn't expose a parseable transcript, so instead of uploading we nudge the agent to call the Rememly checkpoint MCP tool itself. Copy hosts/cursor-hooks.json into your Cursor hooks config and replace /ABS/PATH/TO with the installed package path.

• sessionStart — injects a context block: recall relevant memory first, and the inferred project container (your workspace folder name).
• stop - every REMEMLY_CAPTURE_EVERY firings, returns a followup_message asking the agent to checkpoint the session (decisions, preferences, facts + a one-line diary). De-dup is automatic.

The hook does not provide MCP tools by itself. Connect the host to /mcp or @rememlyai/mcp first; the local package is preferred when the agent should resolve project memory and Todo work from the current git repository.

Codex

Codex reuses the Claude Code capture script. Copy hosts/codex-hooks.json into .codex/hooks.json and fix the path.

Develop

npm install
npm test        # vitest — pure core logic (parsing, cadence, container hints)

The pure, testable core lives in lib/capture-core.mjs; the runners in bin/ are thin I/O shells over it.