| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-11-01 | 1.5 kB | |
| v3.68.0 -- agents remember past sessions (agentdb 3.2 hybrid recall) source code.tar.gz | 2026-11-01 | 71.6 MB | |
| v3.68.0 -- agents remember past sessions (agentdb 3.2 hybrid recall) source code.zip | 2026-11-01 | 77.1 MB | |
| Totals: 3 Items | 148.7 MB | 2 | |
ruflo 3.68.0: agents remember past sessions
Plain-language write-up: https://gist.github.com/ruvnet/7a0f4626da9e22a5bc39fca62dc0e37a
When a session ends with a summary, ruflo now saves it as an agentdb episode, and agents can search those episodes later. The search runs on agentdb 3.2.0 and blends meaning search with keyword search. A paraphrase finds the right session, and so does an exact ticket id or function name. An optional local reranker improves the ordering further.
Added
agentdb_episode-recallMCP tool. It searches past sessions andmemory distillepisodes.rerank: truecosts about 0.1–0.2 s and downloads a 23 MB model on first use.agentdb_session-startacceptsrecall: trueand returns related past episodes.
Fixed
- Session episodes are stored. The bridge called
startEpisode/endEpisode, which agentdb has never had, so nothing was ever saved. Episodes are now stored as neutral (reward 0), because the outcome is unverified.
Changed
- agentdb
^3.2.0, which brings in ruvector 0.3.5.
Measured
- LongMemEval_M, top-5 recall_all (agentdb ADR-016): vector only 0.630 → hybrid 0.719 → with rerank 0.789. The paper's best is 0.762.
- Fresh
npm i ruflo@3.68.0, MCP stdio: - recall in 13–31 ms;
- with rerank, 96 ms once the model is cached;
- stdout is JSON-only;
- one agentdb copy in the tree;
npm audit: 0 critical.
ADR-519 · PR [#4632]
🤖 Generated with RuFlo