Open-source memory tool for coding agents auto-expires stale facts, checks its own answers
is an open-source MCP server that gives , shipped and refined over six months. Unlike plain vector-search memory stores, which kept leaking outdated facts back into , this tool uses a temporal fact graph: every stored fact carries a valid-from and valid-until date, an evidence type (test result, source code, user correction, or session record), and a specific to that evidence type. When two facts conflict, an automatic strategy resolves the disagreement using who confirmed it, how stale it is, confidence gaps, and how many sources back it, scoring 100% (105 out of 105) on the project's own benchmark.
If the evidence falls below its , the system returns no answer rather than guessing, leaving it for manual review. It also runs a 'Coach-Player' adversarial check: an independent Coach LLM cross-checks every material claim in a candidate answer against the supplied source facts and returns a confidence rating (high, medium, or low) plus an itemized breakdown of what was verified versus unverified. On the first hand-labeled hallucination benchmark, it matched exactly on all 12 test pairs (using Claude Haiku 4.5).
Ten runtime adapters are already shipped, covering Claude Code, Cursor, Codex, Continue, Cline, Windsurf, , Hermes Agent, OpenClaw, and pi, connected by registering it as a stdio MCP server.
Key points
- Uses a structured temporal fact graph (valid dates, decay half-lives) instead of plain
- Auto-resolves conflicting facts; defers to manual review rather than guessing when confidence is too low
- An independent Coach LLM cross-checks each claim in an answer against source facts (Coach-Player method)
- Scored 100% (105/105) on its own contradiction-resolution benchmark and 100% (12 pairs) on a hallucination benchmark
- Ships adapters for 10 coding agent runtimes including Claude Code, Cursor, and Codex