What packset measured
packset is the standing-knowledge store in the seat. It writes only on Remember: and Prefer:. The crate explanation has the rows. Neighbouring systems are named by protocol.
What we measured
LongMemEval_S (doi:10.48550/arXiv.2410.10813), 470 answerable questions, session-level retrieval. The pack writes claims, not harvested chat. Source: crates/packset-daemon/examples/longmemeval.rs in packset. Data held privately.
| Arm | hit@1 | recall@5 | recall@10 |
|---|---|---|---|
| packset BM25+ | 0.855 | 0.914 | 0.952 |
| packset fused panel | 0.889 | 0.949 | 0.981 |
Answer accuracy on the same 470 uses a local 7 B reader and judge (Qwen2.5-7B-Instruct Q5_K_M, LongMemEval prompts, top five sessions). CombMNZ fusion is 0.549. Lexical alone is 0.532. Labelled sessions, the ceiling for any retriever, are 0.634. Reproduction: scripts/longmemeval_qa.py and docs/orgmode/results/longmemeval-s-qa.org in packset.
MemoryAgentBench (doi:10.48550/arXiv.2507.05257), reader Qwen2.5-7B. Accurate retrieval scores each source on its own metric, then takes the mean of the four. Lexical job 11759 is 0.568 (95% CI 0.530-0.605, clustered by question). The per-source scores are SH 0.730, MH 0.530, LongMemEval (S*) 0.370 and EventQA 0.640. Published rows are BM25 0.605 and HippoRAG-v2 0.651 with a GPT-4o-mini reader. Conflict resolution at 262K is 0.395 for fused live (job 11750, CI 0.350-0.440) and 0.240 for lexical. The published FactConsolidation rows are BM25 0.255 and HippoRAG-v2 0.295. Hop2 (job 11793) is about 0.555 from the job log, with its verdicts held out. Source: crates/packset-daemon/examples/memoryagentbench.rs in packset. Data held privately.
Neighbouring published numbers
Two public repositories share the name agentmemory. They report different protocols. The first is session retrieval (R@5). The second is question answering with a hosted generator and a hosted judge. packset publishes both a retrieval row and a local-reader QA row so those protocols stay distinct.
The 0.949 fused recall@5 is the neighbour of the 95.2% R@5 retrieval figure. The 96.20% QA figure uses a different generator and a different judge. InjecMEM and the Memory INJection Attack (MINJA) describe the risk of writing from a transcript. packset writes only on Remember: and Prefer:. Privacy-of-write.
What the other habitats answer
Retrieval is one store. The seat also cites frozen product, the tracker, this-session claims, argv law, and weighted agreement. Those questions are not retrieval scores.
| Question | Store |
|---|---|
| What did the human freeze? | cards |
| What does this seat know, standing? | packset |
| What did this unit produce? | deedar |
| What does this node stand on, and who agrees? | vissue |
| What is this session handing out? | claimdag |
| May this argv run? | ljos-policyd |
| Who is right, under trust? | ljos-consensus |
What a default install costs
Measured on 10 October 2026 on one shared box: an 8-thread Intel Xeon, 15 GB of RAM, rustc 1.99.0. Each crate was built from crates.io in a fresh CARGO_HOME and target directory with cargo install --locked -j 2, packset after ljos. The encoder is not part of this install, so search is lexical. Script: scripts/measure-install.sh; raw logs: measurements/2026-10-10-install.
| Measure | Value | Note |
|---|---|---|
| Build ljos 0.27.0 | 170 s | includes the crates.io index and source downloads |
| Build packset 0.12.1 | 90 s | reuses the downloads from the ljos build |
| Binaries on disk | 45.4 MB | ljos 12.2, ljos-mcp 16.0, packset 3.9, packset-mcp 7.5, packsetd 5.8 |
| packsetd resident memory | 4.3 MB idle, 14.7 MB | the second after 281 claims and 330 searches |
| Search, p50 / p95 | 3.3 ms / 4.1 ms | one packset search process per query, lexical, 281 claims, n = 300 |
Two jobs, not eight: the box was short of memory at the time, so a build on an idle machine with every core should be faster. The claims are sentences taken from this site's own pages.
policyd is argv law
ljos policy -- argv prints the line and any pack rule that matches it. ljos rule writes that law into the pack. ljos-policyd is the trusted computing base when it is on PATH or POLICYD_BIN: it prints allow or deny. Reloading a pack is not a check. ljos-policyd docs · Argv field note.
Source: packset explanation · packset