Five agent-memory papers, September 26–October 2: the record stays whole and the decision moves to read time
Five primary-source papers from September 26–October 2 show what changes when a memory system keeps the record whole and decides at read time: what organizational documents can then answer, what raw records with fast judgments score and cost, where evidence is lost between retrieval and the model, who is entitled to see a fact, and what a write-time policy silently discards.
From September 26 through October 2, 2026, five primary-source papers in AI agent memory converged on one relocation. Two of them keep the record whole and move the choice about what matters to read time; the other three locate what that deferral costs, in the form the retrieved evidence arrives in, in who is entitled to receive it, and — for the systems that keep deciding at write time — in the evidence an arrival order quietly discards.
A revised decision is a new document, so nothing is distilled at write time
Date: October 1, 2026
What happened: In an organization, a revised decision arrives as a new document rather than an edit, so a policy set in March, revised in June and reversed in September sits in the record as three documents that all stay. Most memory systems distill each one at write time into facts, notes or graph edges, which fixes what can be answered before anyone knows the question. Mem++ stores every document whole with its date and author, calls no generative model at write time, and defers the choice between versions to read time: retrieval is bounded by the as-of date in the question, lexical and semantic rankings are fused, three of the retrieved slots are reserved for the latest-dated matches, and the answering model sees the surviving versions and chooses. 1
On OrgMemBench — 443 dated artifacts in 157 threads across 18 months, with 73 questions graded against a rubric rather than marked right or wrong — Mem++ takes the best overall score at 57.6 with gpt-4.1-mini, 2.6 points above RAG and 13.1 above A-Mem; with gpt-4o-mini it runs 8.0 points above Zep. Putting the entire history in the prompt scores below 18 overall and near zero on supersession and bi-temporal questions. The limits are stated: RAG still beats Mem++ on contradiction by 28.2 points and on bi-temporal by 16.7, every method stays below 23 on justification chain, and the ablations find that consolidation, extracted-fact cards and an entity graph add cost without a consistent gain. Indexing whole documents beat indexing extracted facts by 2.6 points on LoCoMo even though the answerer read the same documents either way. 1
Why it matters for builders: If your agent answers questions about records written by different people at different times, distillation is the step that decides answerability before you know the question. Store the artifact with its date and author, filter by the as-of date the question carries, and keep superseded versions rather than overwriting them. The write path here costs one index entry per document and no model call, which is the part that survives contact with volume; the compensating work lands in retrieval and in the answering model's judgment about which version holds.
Nothing is decided about a record when it is written
Date: September 28, 2026
What happened: Mnemon keeps conversations as raw, dated records and decides nothing about them at write time. At query time an LLM writes a few search queries, a decision model called Jev answers small yes/no questions about what those searches returned — is this needed, is it no longer current, does it carry the second of two dates — and rules with explicit budgets turn those judgments into a small View that an otherwise unchanged agent receives with the user's message. A background pass consolidates each record once into an index of topic timelines, value histories and standing instructions that points back at the records rather than replacing them. The argument for the design is the data-processing inequality: write-time processing is a fixed function of the history, so it can lose information about questions not yet asked and can add none. 2
Under a public re-evaluation of 14 memory systems with the same answering model, Mnemon scores 91.7% on LoCoMo (next best 88.83%, the rest at most 83.48%) and 83.8% on LongMemEval-S, while sending under 4k tokens of context per question and posting the lowest effective cost index on LoCoMo at 0.259 against 0.337. With a reasoning model answering, it reaches 92.2% and 94.4%. From BEAM-100K to BEAM-10M, 80 times the records, cost per question grows by a factor of 1.11. On the same records Jev separates gold evidence better than two LLMs, at AUC 0.942 against 0.900 and 0.853. 2
Why it matters for builders: The prerequisite is a store that returns dated records from a search route, which most teams already have; memory then arrives without an extraction schema, so a journal, a document set and a task list can feed the same agent. Judge the split on your own workload rather than adopting it wholesale: the fast model does the many small decisions and the LLM does the few slow ones, and that division is where the token budget comes from. Report tokens per question beside accuracy, because the 14-system comparison shows reported scores falling ten to forty points once protocol and grader are held fixed.
Retrieval was perfect and the answer still turned on the pixels
Date: September 26, 2026
What happened: This paper names the stage between retrieval and the answer, calls it delivery, and measures it in isolation: what of the retrieved memory reaches the model, and in what form. On MemLens, feeding the model the gold messages for each question — retrieval at ceiling — and varying only whether their images are delivered moves accuracy by 13.87 points on an 8B backbone, while replacing the retrieved set with the gold set on the same questions and judge moves it 2.31. Delivery is the larger term on all three backbones, growing to 15.61 and 17.92 points, and retrieval grows faster (9.83 and 11.56), so the ratio between them narrows from 6.0× to about 1.6×. 3
DeliverMem turns the finding into three decisions — keep the original modality, give each item a readable identity, and state when it was seen — each measured against a control that alters only its own variable. The rule the controls produce: a delivery decision helps when the information survived storage and reached the model, the question needs the property the decision supplies, and the delivered context does not already carry it; outside those conditions its effect is null. The failure is common enough to see in the field: on DMV-Bench with Qwen2.5-VL-7B, retrieval recall ran 98.4% to 100%, and 98% of the failures picked a neighbouring item while the correct one sat in context, with the agent's own reasoning trace naming the right target in 94% of those cases. 3
Why it matters for builders: An accuracy score folds two stages into one number, and this decomposition splits them: run the same questions with gold context and with retrieved context, and compare the two deltas to learn which stage is costing you. If your agent stores captions in place of the images users sent, the loss shows up at delivery and grows with how fine-grained the question is. Identity and timestamp on each delivered item are cheap, need no training, and were null only where the question did not need them.
The fact was shared with one audience and arrived in another's context
Date: September 28, 2026
What happened: A personal agent acting for its owner across private and shared conversations accumulates one long-term memory, while each fact was shared with the particular participants who were present. This paper extends authorization to the whole memory lifecycle. Every item carries the audience present when it was recorded; a derived item is partitioned by audience, given the intersection of its sources' audiences, or suppressed; an audience widens only through an explicit, object-specific grant; and an item enters a model attempt only when every current viewer belongs to one of its authorized audiences, with unresolved viewers failing closed to public-only. Because exclusion happens before ranking, an inadmissible item never occupies a retrieval slot and cannot displace an admissible one. 4
In a prospectively frozen confirmation over 10,000 multi-party histories, the two reference architectures assembled 200,000 exact contexts and placed zero forbidden items into them, where unscoped retrieval placed 220,385 forbidden items into 82% of its contexts. Entitled recall was non-inferior to a policy-equivalent baseline at a −0.05 Recall@5 margin and ran 0.298 above unscoped retrieval, an advantage that grows with the dose of distractors about other participants. The guarantee holds under stated assumptions: explicit identity, provenance completeness, and complete mediation of every memory path into a model call. The histories come from a controlled event law that fixes participants, audiences and gold evidence before any language is rendered, so the figures describe that generator. 4
Why it matters for builders: This is the confidentiality gap in any assistant that sits in group threads: relevance-only retrieval ignores who is present, output filtering runs after the model has read the context, and per-user silos give up recall the audience is entitled to. Bind the audience from the transport at write time, carry it through every derivation, and check membership at the last step before each model call. The measured cost of doing it is zero recall against an equivalent policy, which makes the trade-off a check on your mediation coverage rather than a budget decision.
Reverse the arrival order and a different record survives
Date: September 28, 2026
What happened: RoutePrism builds memory twice from the same source pool, forward and in exact reverse, holding record text, timestamps, policy, prompt interface and context budget fixed, and records which source identifiers the policy exposed and what compiled context the answerer saw before any answer was generated. A matched four-condition intervention then tests whether the record that reordering displaced carried task-relevant evidence: support present scored 0.873, support absent 0.218, the displaced record restored 0.875, and a length-matched non-supporting replacement 0.296, all sharing question, options, source count, ordering and token length. Restoring that one record recovered 0.657 of lost accuracy, with a 95% user-bootstrap interval of [0.480, 0.808], while the placebo did not. 5
On the primary set of 63 queries from 29 users, every pair changed its exposed source set, answers differed between the two routes in 55.5% of pairs, and 39 of 63 changed the direct-support set as well. Survivor selection — the choice of which record a cluster retains — drives most of the source-level change, and the policies carry distinct signatures: compaction changes which records survive, summarization changes the compiled text while keeping source identities, and A-MEM changes internal metadata without touching the answer prompt. The same restoration effect replicates across five answer models. 5
Why it matters for builders: Endpoint accuracy hides this failure, because a model can answer correctly from wrong evidence and gains on some questions mask losses on others. Build the pair once on your own store: reverse the import order of one real history, diff the exposed sources and compiled context, then restore the displaced record and substitute a length-matched record that supports nothing. Keep the policy that survives that test, and count how many of your questions lose their only supporting record — that stratum is the figure a compaction or recency policy is quietly spending.
Issue summary
Three patterns run across the five items.
- The record is kept whole and the choice is deferred, which moves the cost to a stage that then has to be measured. Mem++ drops the write-time model entirely and pays at retrieval; Mnemon decides nothing at write time and pays in query-time planning and judging; the delivery paper shows the bill for the last stage in the chain, and the authorization paper shows a lifecycle label that costs nothing to carry. Each design pays at a stage the others leave cheap.
- The damage sits in a stratum a builder can count before shipping. Mem++ loses the contradiction and bi-temporal categories to RAG; the delivery gain concentrates on questions that require reading a fact off a particular image; the authorization advantage grows with the dose of distractors about other participants; RoutePrism's recovery holds on the 39 of 63 pairs where the direct-support set changed. Every one of those strata is measurable on a sample before a change goes out.
- The read path's last metre carries effects as large as the store's design. Delivering pixels rather than withholding them was worth 13.87 points against 2.31 for perfect retrieval; restoring one displaced record was worth 65.7 points; skipping a membership check exposed forbidden items in 82% of contexts. Mechanisms at the end of the pipeline deserve the same instrumentation as the store.
One concrete next action
Reversal-test the memory you already run.
- Build the pair. Take one real history from your agent and build memory twice from the same records — normal import order, and its exact reverse — with policy, prompt and context budget fixed. Diff which source records each path exposed and what compiled context the answerer received.
- Restore and substitute. For every question where the exposed sources differ, put the displaced record back, then run the same question once more with a length-matched record that supports nothing. Keep the change only when restoration recovers accuracy and the substitute does not.
- Split retrieval from delivery. Rerun the questions once on gold context and once on retrieved context. If the gold-context delta is large, work on delivery — original modality, item identity, timestamps; if the retrieved-context delta is large, work on retrieval. Then repeat the pair on your next history, because the stratum that lost evidence is the one to watch.
References
- 1
- 2
- 3
- 4
- 5
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.