Mneme — agent memory protocol#

Mneme is a plain-text format for an agent's persistent, cross-session memory. A memory file ( *.mneme) is a list of cells (one decision/fact/preference each, self-contained), and a derived spine (one line per cell: id + state + gist). Full spec: https://github.com/casperkwok/mneme.

Core rule — content is immutable, state is mutable. A cell's gist/body/ cue is a fact about the past, never edited. Its state/ seen/ conf is its relation to now, updated as reality moves. To change content, add a new cell that supersedes the old.

When to use#

• A *.mneme file exists in the project → it is the memory; follow this protocol.
• The user wants durable memory ("remember this across sessions", "记住这个决策/偏好/踩坑").
• You just produced something a future session would be wrong or redo without.

Cell format#

text

@ DEC-0042  storage/cache                 # @ ID(PREFIX-NNNN, unique, never reused) + topic-path
gist  Local cache uses SQLite, not JSON   # one self-contained line; promoted into the spine
state live   conf high   since 2026-06-03   seen 2026-06-03
cue   concurrent writes / cache corruption / local storage choice   # recall keywords
> why / cost / context (free prose, may span multiple > lines)
link  supersedes DEC-0031                 # typed relation, repeatable

• state: proposed (not yet authoritative) · live (current/authoritative) · stale (maybe expired, verify first) · superseded (replaced) · retired (dead, audit only). Spine/working-memory shows only live + stale.
• ID prefixes: DEC decision · OKR goal · PREF preference · GOTCHA pitfall · FACT fact/constraint · TODO long-lived todo. Extensible.
• link relations: supersedes/ superseded-by (pair them), relates, depends-on, blocks, refines.
• Line-oriented, no nesting; unknown lines are ignored; # lines are comments.

Protocol (SOP)#

1. Boot — when project context matters, read only the spine, not whole files: python scripts/mneme_tool.py spine <file> (or grep '^@\|^gist\|^state'). Keep it resident.
2. Recall — before acting, match the task topic against each cell's cue and topic-path; expand only the cells that hit. Recall generously, then judge relevance by content.
3. Trust — live + conf high + recent seen → use directly. stale / conf low / old seen → verify against the real source first, then bump seen to today (raise state/ conf if warranted). proposed → not a conclusion; don't act on it unless the task is to advance it.
4. Write — when something durable is produced, append a new cell: next id via python scripts/mneme_tool.py new-id PREFIX <file> (never hand-count), self-contained gist, set state/ conf, since = seen = today, write cue (which future topics should recall it — don't skip), > for the why, link as needed. Append to end; never silently edit an old cell. Test: would a future session be wrong or redo work without this? If yes, write it.
5. Supersede — to revise a belief: new cell + old cell set superseded + mutual link ( supersedes/ superseded-by). To merely reaffirm: only bump the old cell's seen. Content is never edited in place.
6. Conflict — two live cells disagree → prefer higher conf, then newer seen, and tell the user explicitly; then supersede the loser.
7. Hygiene — one claim per cell; gist self-contained; long content external (link/path only); ids never reused.

Tooling (don't rely on agent discretion)#

scripts/mneme_tool.py (zero deps):

• spine <file> — derive the spine (live/stale).
• lint <file> — health-check: duplicate ids, missing required fields, asymmetric/dangling supersede links, bad dates, and live cells whose seen is older than 45 days (decay). Exit code 1 on errors — wire into a git pre-commit hook so lifecycle metadata can't silently rot.
• new-id PREFIX <file> — allocate the next collision-free id.

Don't use Mneme for#

High-frequency precise queries / transactions (use a DB), large blobs (store externally, link only), throwaway within-session scratch, or things that belong in code/docs. Mneme holds the why — decisions, constraints, lessons, preferences — not every what.