packset

What does this seat know, standing?

A pack of written claims with a review clock, not a transcript with an index.

Remember / Prefer only LMDB, one writer CLI + HTTP + MCP

A pack is what a seat has decided to keep: one claim per atom, written on purpose, never mined from a transcript. Search ranks the claims by a measured panel of scorers. A review clock brings each claim back before it is forgotten, and a claim that is not reviewed decays in the ranking. The writer is one daemon over one Lightning Memory-Mapped Database (LMDB) file, and every other tool is a client.

Install#

$ cargo binstall --locked packset          # packset, packsetd and packset-mcp
$ packset ensure

First minute#

$ packset remember --workspace demo "The lexical default is BM25+. It scored 0.635 hit@1 against 0.615 for BM25."
ee17d2a07cbf1569e54e0d80496fdecf    lesson  due 2026-09-28T16:52:09.119Z
$ packset prefer --workspace demo "CombMNZ over RRF for fusing two ballots."
63df91bbea2b5e4e146866db0a85b4c6    preference  due 2026-09-28T16:52:09.149Z
$ packset search --workspace demo which fusion
packset: workspace demo
2.0000  preference  63df91bbea2b5e4e146866db0a85b4c6    CombMNZ over RRF for fusing two ballots.
1.0000  lesson  ee17d2a07cbf1569e54e0d80496fdecf    The lexical default is BM25+. It scored 0.635 hit@1 against 0.615 for BM25.

Two claims, written on purpose, each with a review date a day out. A search lists the standing choice first, then the lesson. The tutorial goes on from these two claims to review and handover.

packset ensure starts the writer when it is down. With nothing set, every client uses that writer on the default port. The dense scorer is a separate binary, packset-embed, built against Open Neural Network Exchange (ONNX) Runtime compiled from source. The pack answers without it, and a sitting’s recall stays lexical.

Measured#

LongMemEval_S, 470 answerable questions, session-level retrieval. BM25+ alone is 0.914 recall@5. CombMNZ of BM25+ and multilingual-e5-large is 0.949 recall@5 (0.889 hit@1, 0.981 recall at 10). QA on the same 470 with a local 7B reader is 0.549 against a labelled-session ceiling of 0.634. Writes only on Remember: and Prefer:. The explanation is the source for those rows.