Five agent-memory papers, October 17–23: the raw record won — a naive retriever first in six of seven settings, and a build that needs 372 questions to pay back
Five primary-source papers from October 17–23, 2026 show the part of an agent-memory stack closest to the raw record carrying the measured gain, while the extraction, abstraction and distillation around it carried the bill.
From October 17 through 23, 2026, five papers in AI agent memory took a published gain apart to ask which component produced it. A dense retriever over raw conversation history finished first in six of seven benchmark-model settings.
An agent that stored the rules it inferred from its own runs scored below one that stored only the runs. Distilling a memory-augmented teacher into model weights bought 1 point; attaching the teacher's retrieval bought 34.
Memory for how a user wants to be assisted performed at the level of no memory at all, unless it was kept as files. And the build cost of a memory system took 372 questions on one history to earn back.
Four of the findings share a shape: what kept the record intact carried the result, and what transformed it carried the bill. The fifth prices the whole arrangement against a stronger reader model.
The items run in date order, each with its numbers, its boundary, and the check a builder can run.
Storing the rules an agent inferred from its own runs kept its errors alive
Date: October 19, 2026
What happened: Ruhan Wang, Kishan Panaganti and Dongruo Zhou, at Indiana University and Tencent HY, describe error reproduction. A faulty stored abstraction shapes a later execution, and that execution becomes the raw material for the next abstraction. 1
They compare four memory designs on ALFWorld, WebShop and ScienceWorld, with three backbones and ten memory-update rounds. 1
Their own design, Anti-Persistence, keeps trajectories and outcomes, synthesizes guidance when a task arrives, and discards that guidance after execution. 1
Evaluation freezes the store at the last update, so each held-out task measures what an agent reuses rather than what it relearns. 1
Anti-Persistence had the highest held-out success in 7 of 9 backbone-benchmark settings. 1
On ALFWorld with Qwen3-Coder-Next it scored 96.0% against 84.4% for trajectory-only memory. On ScienceWorld with Llama-4-Scout it scored 30.6% against 20.0%. 1
Storing trajectories and abstractions together did worst on Llama-4-Scout's ALFWorld, scoring 33.3% against 41.6% for trajectories alone. Anti-Persistence also used the fewest model calls per task of the four designs. 1
The evidence for the error-reproduction loop is an association between earlier and later generated abstractions, which the authors report without isolating it through an intervention. 1
Why it matters for builders: If you persist skills or summaries, your store becomes the training set for your own mistakes. Keep the trajectory and the outcome, generate guidance for the task in front of the agent, and throw the guidance away.
Where you do keep an abstraction, keep the run that produced it, so the abstraction can be re-derived rather than inherited.
Distilling memory into weights bought one point; the same memory attached at inference bought 34
Date: October 21, 2026
What happened: Cheng-Yin Tu, at Nikon Research Corporation of America, pre-registered the hypotheses, arms, tests and analysis rules before training began. 2
A vision-language teacher retrieves lessons, with episode images attached, from its own past attempts at a single-arm reach task. LoRA student policies are then distilled from the trajectories that teacher produced. 2
Distillation beat an action-only control by 1 percentage point, 26 of 100 against 25 of 100, with a 95% interval from -11 to +13. 2
That pre-registered test was underpowered for its own minimum effect of 10 points, and the authors withdraw the memory-in-weights claim rather than refute it. 2
Attaching the teacher's frozen retrieval buffer to fixed student weights added 34 points (z = 5.15) and reached the teacher's pooled reference of 49.3%. 2
The pre-registered protocol contrast agreed, at +23 points (z = 3.36). In further controls registered before they ran, the same buffer lifted the base model with no adapter at all to 48%, and three other adapters from 19-26% to 53-55%. 2
On a second, harder task the retrieval advantage pointed the same way and fell short of significance: +6 points, 95% interval from -4.5 to +16.4. The paper's diagnostics tie that attenuation to how much successful experience the buffer had to draw on. 2
Why it matters for builders: The authors read the result as distillation carrying the marginal correction that every arm adopts, and losing the conditional, situation-specific correction that retrieval supplies. 2
When your experience is conditional, keep it where the model can address it at inference, and test the harder task before assuming the gain transfers.
For interaction preferences, retrieval memory did not beat having no memory
Date: October 21, 2026
What happened: Jialin Li, Zhenhao Chen, Hanjun Luo and Hanan Salam, at New York University Abu Dhabi, release MemPABench. Four simulated users carry 14 interaction attributes across work and personal contexts, over 99 sessions and four months, with five preference changes per user. 3
The environment runs 83 executable tools across 26 domains, and a judge blind to the agent's own selection scores each reply against the target setting. 3
Four memory systems run on four backbones against two references: no memory at all, and an oracle that hands the agent the correct preference. 3
File memory was the only condition that clearly improved preference selection, by 9.6 points, with a 95% interval from +3.9 to +15.6. It beat no memory in 14 of 16 user-backbone pairs. 3
Mem0, MemOS and simple retrieval moved by +0.4, +0.4 and +0.2 points, with intervals inside 2.7 points of no memory. No memory beat them in 10, 8 and 8 of the 16 pairs. 3
File memory also carried preferences into the context where they were wrong: 58 carry-over instances against 13 for no memory, out of 216. 3
The oracle exposes a second gap. Handing the agent the target preference raised correct selection of the five hardest settings from 33% to 96%, while replies rated 4 or higher on the paper's behaviour rubric moved only from 24% to 59%. 3
Under no memory, preferences about what the agent does were selected correctly 40.3% of the time, below the 44.4% of random choice. 3
Why it matters for builders: Selecting a preference and acting on it are separate measurements, so score the reply and the action the agent took. Store each preference with the context it applies to, because a store that remembers well will also apply what it remembers where it does not belong.
The build earns itself back after 372 questions, and less as the reader gets stronger
Date: October 21, 2026
What happened: Zedong Peng, at Shanghai Jiao Tong University, costed four memory methods and three coding agents on one LoCoMo conversation of 419 turns and 152 questions, across three models. 4
With GPT-4o-mini, Mem0 led at 78% against 74% for full context and 74% for BM25, while Zep scored 43%. 4
Mem0's build alone cost $0.56 against $0.43 for one full-context rollout on that model. On Claude Opus 5.5 the same pair ran $25.88 against $17.84. 4
The paper puts break-even at 372 questions for GPT-4o-mini, 545 for GPT-6 Luna and 596 for Opus 5.5, on the same history. 4
With the strongest model the ordering changed. The Pi coding agent scored 99% at $6.95 per rollout, against 95% for both full context and Mem0 at $17.84. 4
Pi read about 10,000 input tokens per question, against 28,900 for full context. 4
One confound cuts the lead. The agents received the LoCoMo file carrying captions for shared photos, and the memory methods received text only. On the 88 questions whose evidence sits in no photo, Pi and full context differ by one answer. 4
The scope is one conversation and one run per cell, which the paper states. 4
Why it matters for builders: Price the ingestion and count the questions it needs to earn back. A tool-using agent that reads raw history is a memory baseline in its own right, and its win has to survive an input-parity check: six of the seven questions Pi answered where full context failed had their evidence in a turn that shared a photo.
A plain retriever over raw history finished first in six of seven settings
Date: October 22, 2026
What happened: Zhengye Bao and colleagues test Raw-RAG: 256-token windows with 32-token overlap, 128 dense candidates, and the top 15 reranked chunks expanded to complete turns. 5
Across LoCoMo, LongMemEval, MEME and MemoryAgentBench, Raw-RAG placed first among the compared memory systems in six of seven benchmark-model settings, ahead of the strongest by 0.13 to 5.53 points. 5
The margins were 0.65-0.95 points on LoCoMo, 0.13 on LongMemEval, 2.24 on MEME with GPT-4o-mini, and 4.48-5.53 on MemoryAgentBench. 5
Coverage explains the result. Raw-RAG retrieved sufficient evidence for 45.29% of MemoryAgentBench questions and 76.35% of MEME questions, the highest on both. Utilization across the seven systems spanned a narrower 80.00-89.71% on MemoryAgentBench. 5
On MEME after-state questions, 62.34% of the questions Raw-RAG covered also carried the complete chain of order or dependency. MemMachine reached 56.70% and MemU 9.62%. 5
The arm is simple, and its own parts still matter: removing the reranker costs 3.92 points, and returning one chunk instead of fifteen costs 11.73. 5
Adding memory management to that baseline cost accuracy on MemoryAgentBench's 14-source total of 48.32%. Atomic-fact extraction dropped it to 28.70%, summary augmentation to 42.51% and graph expansion to 42.54%. 5
The additions also cost time and tokens: 21.6-109.3 million language-model tokens and 2.9-8.8 hours to ingest, against 12 minutes for Raw-RAG. 5
Two boundaries hold. MemMachine led on MEME with DeepSeek-V4-Flash by 1.52 points, and full context led Raw-RAG there by 19.98. On PersonaMem-v2, Raw-RAG placed third of seven at 49.62%, behind MemOS and EverMemOS. 5
Why it matters for builders: Run the raw-history retriever first, with the reader, prompt and evidence budget you already ship, and treat any ingestion step as unproven until it beats that arm. Track coverage and utilization as separate numbers, because the systems here differed far more in supplying the evidence than in using it.
Issue summary
Three patterns run across the five items.
- The record carried the result. Raw chunks beat extracted facts, file memory was the only store that moved preference selection, retained trajectories beat retained abstractions, and a retrieval buffer beat a distilled policy. The artifact closest to the original record did the work in each case.
- The bill went to the transformed form. Fact extraction added 21.6-109.3 million tokens and up to 8.8 hours of ingestion, a Mem0 build needed hundreds of questions to earn back, and storing abstractions alongside trajectories produced the lowest scores in the error-reproduction comparison.
- Knowing was the smaller half. An oracle preference raised selection to 96% and moved delivered behaviour to 59%, a stronger reader model closed most of the gap the memory was meant to close, and a distilled policy kept the static prior while dropping the conditional one.
One concrete next action
Build the naive arm first, on your own workload.
- Assemble the raw baseline. Chunk your history at a few hundred tokens, embed it, rerank the candidates, and expand the winners to complete turns. Use the reader model, prompt and evidence budget you already ship.
- Swap representation with the evidence held fixed. Add fact extraction, summaries or graph expansion to that arm one at a time, and record coverage alongside the accuracy conditional on coverage. Falling coverage with steady utilization points at relations the transformation dropped.
- Price the ingestion. Record build cost and the question count at which it earns back, then repeat the comparison with the strongest model you would deploy. The reader sets the ceiling on what a store can return.
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.