Agent memory stops hoarding transcripts: distilled claims with evidence still land in LongMemEval's top ten 99.8% of the time

The unit of agent memory changes from raw conversation chunks to claims: each claim records who stated what, when it was true, over what period, and which span of the source proves it, while everything else is discarded on purpose. The common approach keeps everything, embeds it, and pulls the nearest chunks into the prompt, asking the model to fill the gaps between fragments. On LongMemEval, a public benchmark of 500 questions over long chat histories, the right evidence still landed in the top ten results on 99.8% of questions.

Memory writes pass through a single front door that can refuse a submission, rather than being indexed first and repaired later. A plain store indexes everything it receives, so bad data enters the retrieval pool alongside good data. In one benchmark run a conversion bug began emitting human-readable dates instead of machine-readable timestamps; the door quarantined 5,732 submissions and accepted zero.

Time and absence become explicit: a new fact supersedes an old one while both stay on the record with their dates, and every question resolves to answered, searched and found nothing, or outside what the system knows. Under append-and-retrieve, an old chunk and a new chunk about the same fact are just two chunks, and the higher-ranked one wins; an empty retrieval also leaves the model unable to tell nothing found from not tracked. With the same system and data, instructing it never to say information is missing moved the overall score from 88.0 to 87.2, but turned 30 honest refusals into 30 fabricated answers; scored only on answerable questions, the same answers read 92.8.

Permissions are decided before the model runs: identity is resolved first, retrieval happens under that person's authorization, and only then is a model called, with disclosure levels set when a fact is written rather than redacted at read time. Most teams either share one context among everyone or tell the model in the prompt what it must not say, which leaves the model enforcing the policy. In the report a finance reader asking about a revenue figure gets 2,140,000, while a sales reader asking about the same subject gets a band of 2.0M to 2.5M and an up trend, with no byte of the precise value in their evidence; across five pairs of isolated memory stores, all 45 adversarial probes were refused or contained.

The demonstrated setting is organizational agent memory across three public benchmarks, one harness, and a public report with per-question results. The author states the next step is turning repeated shapes into standard procedures and then into narrow purpose-built agents or plain code.

The distillation is explicitly described as lossy, and rebuilding the record with Gemma 4 moved answer quality by one to two points, so the choice of extractor affects results. The author also notes that scoring only answerable questions makes the same answers read higher.