Without a lifecycle, a memory file accumulates the January setting and the March override and hands the model both. The status ladder plus supersede links is a direct answer to that, and the conflict and staleness passes turn the problem into something you can inspect rather than discover by being given a wrong answer. The author reports tuning the governance engine down from 1,848 false positives to 11 real findings, which is the sort of number that suggests the naive version does not work.
A local memory server whose distinguishing idea is a status on every memory. Most memory tools solve storage; this one is aimed at trust — recall ranks validated memories above active ones above hypotheses, and excludes deprecated and superseded ones entirely, so an outdated rule cannot be retrieved alongside the rule that replaced it. Storage is SQLite with WAL mode, search is FTS5, and there are no API keys or cloud calls.
- 13 MCP tools covering the whole loop: `add_memory`, `search_memory`, `recall_memory`, `update_memory`, `delete_memory`, `list_memories`, `save_session`, `load_session`, `promote_memory`, `deprecate_memory`, `supersede_memory`, `memory_health`, `memory_conflicts`
- A status lifecycle you drive explicitly — promote a hypothesis to active and then validated, deprecate what is no longer true (kept for history, excluded from recall), or supersede it with a newer memory that the old one now points to
- Conflict detection that separates duplicates from real contradictions using topic overlap and sentence-level negation matching, with `critical` severity when both sides are active and `warning` when one is deprecated
- Staleness detection by age, by a referenced source file having disappeared, and by hash drift — the source file changed but the memory did not
- A health score out of 100 built from conflicts, stale percentage, orphaned references, deprecated weight, and whether you have any validated memories at all
- Context-budgeted recall that packs the most trustworthy relevant memories into a token limit, scored on relevance, trust status, provenance, recency, frequency and confidence
- Seven memory types that match real agent work: `setting`, `bug`, `decision`, `procedure`, `context`, `feedback`, `session`
No account, no API key, no cloud, no vector database — just a `.db` file. `pip install quilmem[mcp]`, then `agentmem init --tool claude --project myapp` and restart your editor; `memory_health` confirms it is live. The PyPI package is `quilmem` and the command is `agentmem`, launched as `agentmem --db ./memory.db --project myproject serve` over stdio. Python-only use works with plain `pip install quilmem`, without the MCP extra. The repository also ships agent instructions to paste into `CLAUDE.md`, `.cursorrules` or `AGENTS.md` so the agent knows when to search versus add. MIT licensed.
One command — pip install quilmem[mcp]