Skip to content
Documentation
Plugins
memory

memory

One memory instead of two, with zero LLM cost: notes the agent writes, SQLite FTS5 search, progressive disclosure.

SurfacesMCP mem_save, mem_search, mem_get; PreCompact checkpoint; SessionStart recall
Specthe spec (replaces) column of the catalogue in plan.md §1
Defaulton

Tools

  • mem_save(kind, title, body, project?) — project defaults to the git root name of cwd.
  • mem_search(query, limit=5) — ids, titles, 120-char snippets ranked by FTS5 bm25.
  • mem_get(id) — full body.

Hooks

  • PreCompact: extracts the last 20 user prompts (≤ 300 chars each), touched file paths and last error lines from the transcript into a checkpoint note.
  • SessionStart with source == "compact" (and PostCompact): injects the latest checkpoint (≤ 400 tokens) through inject.
  • SessionStart recall: last 5 note titles + ids for the project (≤ 200 tokens, priority 10), never bodies.

Import

rtok memory import <file.jsonl> reads one note per line ({kind, title, body, ts?, project?}), deduped by body sha256. Export your previous memory tool to that shape yourself; rtok knows no third-party schema (D6).

Tasks

See roadmap.md § memory. Checks in plan.md.

T6.1 notes API · T2.5 checkpoint · T6.2 recall · T6.3 import.

Status

Manifest only. Schema (notes, notes_fts) exists since T0.3.

measureproxy