Can memory-guided rereading improve agentic long-context reasoning?

Memory overwriting discards evidence before it looks useful

Streaming reading agents avoid the quadratic cost of standard attention by reading a long document in chunks and folding each chunk into a single evolving memory. The paper argues that this design has a structural weakness: information that looks unimportant when first read can be overwritten before a later bridge fact reveals its relevance. Once that evidence has left the memory, it cannot be recovered.

The paper's case comparison on 2WikiMultiHopQA shows the failure concretely. One baseline, MemAgent, identifies the film's director but compresses the birthplace fact into the coarse summary "born in Germany"; when the precise birthplace Elberfeld later appears, the memory no longer holds a specific enough anchor to capture it. A second baseline, ReMemR1, abandons an entity-resolution query too early, so the birthplace passage is bypassed and later retrieval has nothing to find.

The paper names two consequences of adding retrieval to this pipeline. Permanent latent evidence loss means discarded evidence never entered historical memory and therefore cannot be retrieved. Interference from invalid queries means the agent cannot tell whether missing information was overwritten, is still unread, or is absent, so premature queries pull in irrelevant content that contaminates memory.

Rereading replaces intermediate retrieval with targeted passes

MemReread keeps the first streaming pass unchanged: the model reads every chunk in order and produces a final memory. The difference comes after that pass. If the memory is judged insufficient, a decomposition step turns the original question plus the current memory into a sub-question, and the agent rereads the document specifically for that sub-question, producing a sub-answer that is integrated back into the root memory.

This is why the paper describes the method as circumventing intermediate retrieval. Rather than searching a store of past memory states, the agent goes back to the source chunks with a sharper question. The paper argues this recovers indirect evidence that was prematurely discarded while preserving the logical flow of reading the document in order.

The loop is bounded by a rereading-pass limit p. Each pass can add one sub-question and answer to a history list, and the process stops early when the decomposition step produces no further question. The final answer is then read out from the updated root memory. In the paper's worked example, the agent first resolves a mother's identity, integrates it, then rereads for her death date and integrates that, ending with a memory that preserves both supporting facts.

Reinforcement learning decides how many passes to spend

A fixed rereading budget would waste computation on easy questions and underspend on hard ones. The paper introduces Rereading-Adaptive GRPO, a reinforcement learning strategy that adapts the number of rereading passes to task complexity, which the authors present as flexible control of computational overhead.

The adaptation is visible in the reported averages. Without reinforcement learning, MemReread-4B averaged 2.03 rereading passes on RULER-QA, 2.50 on LongBench-E-QA and 2.61 on LongBench-v2. With ReA-GRPO those averages became 0.17, 2.91 and 2.24 respectively, while scores rose from 54.8 to 64.0, 47.2 to 50.9, and 27.1 to 32.1 percent. The near-zero average on RULER-QA shows the policy can collapse rereading almost entirely on that benchmark rather than spending a uniform budget.

The paper also reports that increasing the maximum number of rereading passes consistently improves performance in ablation studies, so the adaptive policy operates within a ceiling that still matters. The authors further state that the reinforcement learning framework improves length extrapolation, which is what allows the same procedure to be applied at context lengths beyond those seen in training.

Gains appear on multi-hop and long-context benchmarks

The headline comparison is on 2WikiMultiHopQA, a multi-hop question answering benchmark. MemReread-4B outperforms strong baselines there, including a +1.6 percentage point advantage over DeepSeek-V4-Pro. The paper's scale comparison table shows the same pattern across model sizes: at 4B, 8B, and above 200B and 1000B parameter backbones, MemReread with a single rereading pass (p_c = 1) has the highest average accuracy in its group, reaching 73.1 for Qwen-Plus and 73.9 for Qwen-Max.

On RULER-QA at the 4B scale, MemReread averages 63.9 against 62.6 for ReMemR1 and 55.8 for MemAgent, and it holds 60.4 at 1M context where ReMemR1 reaches 54.2. On LongBench-v2-1M it averages 32.1 versus 28.3 for MemAgent and 25.0 for ReMemR1, with the largest gap on the Long split (38.0 versus 23.1).

The LongBench-QA and LongBench-E-QA results are more mixed and worth separating from the headline. On LongBench-QA, MemReread averages 44.9, slightly above MemAgent's 44.6 and ReMemR1's 44.1, but it is lower than MemAgent on 2wikimqa (67.2 versus 65.2 in MemAgent's favor) and on dureader. On LongBench-E-QA it averages 50.9 against 50.4 for MemAgent, with its advantage concentrated in the 8k+ bucket (51.1 versus 47.0). These are different benchmarks with different denominators, so the multi-hop gain should not be read as a uniform improvement everywhere.

The mechanism transfers across backbones and checkpoints

The paper tests whether rereading depends on a particular training recipe. Initializing MemReread with MemAgent weights yields substantial gains even without Rereading-Adaptive RL training, and initializing with ReMemR1 weights gives comparable performance despite operating without any explicit retrieval module and using entirely different prompts. The authors read this as strong transferability of the framework.

Zero-shot results across backbones point the same way. With p_c = 1 and no training, MemReread averages 71.6 on Deepseek-V4-flash, 73.1 on Doubao-Seed2.0-lite, 66.3 on Gemini-2.5-flash and 75.0 on GPT-4.1-mini on 2WikiMultiHopQA, each above the corresponding MemAgent and ReMemR1 averages in the same table. The authors interpret these findings cautiously, suggesting that rereading guided by memory-based question decomposition may function as an intrinsic reasoning mechanism in streaming agents provided the model retains sufficient contextual information, and they leave rigorous characterization to future work.

Most remaining errors come from misreading during updates

The paper's failure analysis locates MemReread's errors in the Read and Integrate stages of memory updating rather than in the rereading trigger. Reading misinterpretation accounts for 90% of sampled reasoning failures: the model misreads a context segment, so the updated memory drifts from the underlying evidence and the final memory is inaccurate. Ineffective integration accounts for the other 10%: the model struggles to reconcile conflicting information and may discard valid updates or let incorrect information corrupt the root memory.

The authors attribute both failure modes to limitations in the backbone model's intrinsic reasoning capabilities. That matters for interpreting the benchmark numbers: the rereading loop can only recover evidence the model is able to read and integrate correctly when it goes back.

The paper also scopes its comparison deliberately. It focuses on methods that keep a context window below 8K tokens at every stage, and it argues that chunk-level retrieval requires storing the entire document, struggles to integrate evidence scattered across chunks, and incurs high peak computation when multiple chunks are processed jointly. The reported advantages therefore apply to that memory-based, bounded-window setting rather than to every long-context architecture.

Rereading pays most when evidence is indirect and scattered

The paper's own comparison splits the decision. Where a task requires resolving an alias or entity before a later fact becomes usable, as in the compositional bridge cases and the global reasoning counting tasks, the baselines fail because the bridging evidence is discarded or noisy retrieval corrupts the count, and rereading with a sub-question is what recovers it. Where the required fact is direct and the first pass already captures it, the paper's own LongBench-QA table shows MemAgent matching or slightly exceeding MemReread on individual datasets.

The practical threshold implied by the evidence is therefore the shape of the evidence chain, not context length alone. MemReread's advantage grows on multi-hop and 1M-token settings where the answer depends on facts that only become relevant after an intermediate resolution, and it narrows on benchmarks where a single pass already surfaces the needed fact. The adaptive pass count is the paper's answer to the cost side of that trade-off: it spends rereading where complexity justifies it and nearly eliminates it where it does not, as the 0.17 average on RULER-QA shows.

The most useful reading of this paper is a comparison between two failure routes rather than a claim that rereading always wins. Retrieval-based agents lose evidence that was never archived or contaminate memory with invalid queries; MemReread instead risks misreading during the reread and integration steps, which the paper traces to backbone reasoning limits. Which risk dominates is the question a reader should carry into a new long-context task.

Sources used in this answer