MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading

arXiv:2605.10268 · cs.CL, cs.AI · Submitted 2026-05-11 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading".

Jane: The gist The proposed MemReread framework enhances streaming reading by recovering overlooked critical facts through an adaptive rereading mechanism,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're looking at this paper titled "MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading." It’s tackling the problem of long context reasoning where standard attention gets quadratically complex, so they’ve moved towards agent memory systems that process chunks sequentially.

Jane: Right, and the core idea here is that these agents keep a dynamic memory while reading the document in pieces. But existing methods have a couple of big problems—they rely on retrieval to get stuff back if it gets overwritten, but that retrieval itself can cause evidence loss or interference from bad queries.

Lu: So MemReread tries to fix this by ditching intermediate retrieval entirely and instead triggering question decomposition and targeted rereading when the final memory isn't enough. This lets them recover indirect facts that might have been prematurely discarded because they aren't in the final memory state m!

Meng: That sounds like a way to keep it linear in terms of context length, which is important for practical use. What’s the big claim they are making about this new approach?

Tom: They claim that by using this rereading mechanism, MemReread enables non-linear reasoning while still preserving the logical flow of how you read a document. It’s all about recovering those facts that were lost during memory formation.

Jane: Basically, they show how to bypass the retrieval problem by focusing on rereading with sub-questions instead of trying to retrieve what was already discarded or overwritten, which helps with recovery without interference.

Lu: And they introduce a reinforcement learning framework called Rereading-Adaptive GRPO that helps them determine exactly how many rereading passes are needed based on how complex the task is. This lets them control the computational overhead flexibly.

Meng: That adaptive part is what I like; it means they aren't just blindly rereading everything, but only when necessary, which makes sense from an engineering standpoint for managing runtime.

Tom: And on the results side, their MemReread-4B model outperforms strong baselines on long-context reasoning tasks, specifically showing a plus one point six percentage points over DeepSeek-V4-Pro on the 2WikiMultiHopQA dataset <ref:2605.10268#pg3>.

Jane: That performance gain is solid, and they keep the time complexity linear with respect to context length, which is a significant achievement given what they are trying to do.

Lu: The analysis shows that this framework successfully preserves indirect facts, like when it recovers the indirect fact director → Peter Paul Brauer and translates it into a precise second-hop sub-query: What is the birthplace of Peter Paul Brauer?, leading to the correct final answer.

Meng: I see how that works in practice, by turning a big question into smaller steps during rereading, which seems like a very logical way to handle complex reasoning chains.

Tom: They also showed that their memory storage overhead is comparable to MemAgent because they don't need to store historical memory at every single step, which keeps things efficient.

Jane: So what’s the main thing we should take away from "MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading"? It’s about using structured rereading guided by decomposition to recover critical facts without resorting to external retrieval modules.

Lu: The implication is that rereading guided by memory-based question decomposition could function as an intrinsic reasoning mechanism within streaming agents themselves.

Meng: From a practical standpoint, this means we might be able to build more robust long context agents that can handle complex dependencies better than what we have now with simple memorization.

Tom: So while the methods are complicated, they’ve managed to maintain linear time complexity while achieving better accuracy on those challenging long-context reasoning tasks.

Jane: That's the gist of it—it improves agentic reasoning by intelligently rereading when needed, which helps them handle complex questions without exploding in complexity like standard attention mechanisms do.

Conclusion: Tom: So we’re wrapping up on MemReread, this paper by some folks that’s trying to tackle those long context problems with sequential reading and adaptive rereading.

Jane: Right, so they put this whole streaming reading thing together with a memory buffer and a way to trigger more questions when the final memory isn't enough.

Lu: The title itself is pretty descriptive, focusing on how memory guides the rereading process for agentic reasoning in long context scenarios.

Meng: It’s about recovering critical facts that might have been missed during the initial pass by systematically rereading with targeted sub-questions.

Lalam: And they use a reinforcement learning setup, called ReA-GRPO, to figure out how many times to reread based on how complex the task is.

Tom: That adaptive part is key because it means the system isn't just blindly rereading everything; it’s smart about when to stop and when to dig deeper.

Jane: It really moves beyond just dumping massive amounts of text into a model and trying to make it do everything at once in one go.

Lu: I think what they show is that this structured rereading acts like an intrinsic reasoning mechanism for these streaming agents, not just an external retrieval hack.

Meng: From a practical standpoint, the fact that it maintains linear time complexity with respect to context length is a big deal for how we build these things at scale.

Tom: So the main implication is that we can get better performance on those tricky long context tasks while keeping the computational cost manageable.

Jane: It suggests a way to handle complex document comprehension without hitting that quadratic complexity wall so hard.

Lu: It opens up possibilities for agents to reason more deeply over long documents in a step-by-step, guided fashion.

Soochow University · Peking University

cs.CL, cs.AI

Submitted: 2026-05-11

Updated: 2026-10-09

Code: https://github.com/iiGray/MemReread

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: The gist The proposed MemReread framework enhances streaming reading by recovering overlooked critical facts through an adaptive rereading mechanism, which supports non-linear reasoning while

Key concepts

Streaming Reading: This is the core method where an agent processes a long document chunk by chunk sequentially, maintaining a bounded memory buffer as it ingests information. It avoids the quadratic complexity of standard attention mechanisms by processing text in a linear fashion, making it efficient for very long contexts.
Memory-Guided Rereading: When the final memory state is insufficient to answer a question, MemReread generates a sub-question and triggers targeted rereading of the document. This mechanism recovers indirect evidence that might have been prematurely discarded during initial processing, enabling non-linear reasoning.
Rereading-Adaptive GRPO (ReA-GRPO): This is a reinforcement learning framework used to optimize MemReread. It dynamically determines how many times the system should reread the document based on the complexity of the task. This adaptive control helps manage computational overhead while maximizing performance.
Linear Time Complexity: This means that as the length of the input context increases, MemReread's processing time grows linearly, not quadratically. This efficiency is crucial for handling very long documents without incurring prohibitive computational costs associated with standard attention models.

Terminology

Agent Memory — 1× in this paper
Attention Mechanism — 1× in this paper · explained in 3 episodes
Computational overhead — 1× in this paper · explained in 2 episodes
Reinforcement Learning — 1× in this paper · explained in 4 episodes

Summary

The gist The proposed MemReread framework enhances streaming reading by recovering overlooked critical facts through an adaptive rereading mechanism, which supports non-linear reasoning while maintaining linear time complexity with respect to context length.

Limitations of Existing Approaches

Retrieval-based approaches suffer from both evidence loss during memory formation and interference induced by invalid queries, as retrieval cannot recover facts discarded from chunks and agents fail to discern whether missing information is overwritten or unread, triggering ineffective queries. Memory-only approaches are constrained by inherent sequentiality, which hinders revisiting overlooked facts and compromises the ability to recover indirect facts that were prematurely discarded.

MemReread Workflow

MemReread operates in four phases: Read, Decompose, Integrate and Answer. During the reading phase, the system adopts the MemAgent paradigm by maintaining a bounded memory buffer throughout ingestion. Upon completing a reading pass, it prompts the model to identify missing information from the final memory state m, and if gaps exist, it generates a sub-question q and performs streaming reading guided by q. The integration stage updates the terminal memory solely with the sub-question and its corresponding answer, directly discarding the intermediate sub-memory.

Training Strategy

To optimize the architecture, MemReread introduces a reinforcement learning framework called Rereading-Adaptive GRPO (ReA-GRPO) that enhances length extrapolation capability while dynamically determining the number of rereading passes based on task complexity. The training objective combines process supervision and outcome supervision through an advantage calculation defined in Equation 3. This objective encourages minimizing the rereading passes without compromising overall performance, as shown by the Rereading-Adaptive Outcome Advantage which follows principles inspired by DRPO and GRPO.

Experimental Validation

Extensive experiments demonstrate that MemReread consistently outperforms baseline frameworks on long-context reasoning tasks, while maintaining linear time complexity with respect to context length. The method achieves superior task performance while requiring 3–4× the average test time compared to MemAgent on 2WikiMultihHopQA. Furthermore, its memory storage overhead is comparable to MemAgent, as it does not require storing historical memory at each step.

Key Findings and Analysis

The framework's effectiveness is demonstrated across various benchmarks, with the 4B model achieving up to 12.1% higher accuracy over ReMemR1 on out-of-distribution datasets. The analysis of failure cases shows that MemReread successfully preserves the indirect fact director → Peter Paul Brauer and seamlessly translates it into the precise second-hop sub-query What is the birthplace of Peter Paul Brauer?, leading to a correct final answer. However, it also exhibits limitations, such as Inherent Reasoning Deficit where it focuses on an incorrect entity during the first integration.

Conclusion

In this work, we propose MemReread, a novel framework that enhances the streaming reading paradigm by recovering overlooked critical facts through an adaptive rereading mechanism. To optimize this framework, we introduce Rereading-Adaptive GRPO (ReA-GRPO), a context-aware reinforcement learning strategy that adaptively modulates rereading passes based on task complexity. Empirical results demonstrate the superior performance over baselines, and ReA-GRPO effectively equips it with task-adaptive rereading capabilities. Ablation studies further confirm that increasing the maximum number of rereading passes consistently enhances overall performance. The work suggests that rereading guided by memory-based question decomposition may function as an intrinsic reasoning mechanism in streaming agents.

The gist

MemReread enhances streaming reading by recovering overlooked critical facts through an adaptive rereading mechanism, which supports non-linear reasoning while maintaining linear time complexity with respect to context length.

How it works

Limitations of Existing Approaches: Retrieval-based approaches suffer from both evidence loss during memory formation and interference induced by invalid queries.
MemReread Workflow: MemReread operates in four phases: Read, Decompose, Integrate and Answer.
Training Strategy: To optimize the architecture, MemReread introduces a reinforcement learning framework called Rereading-Adaptive GRPO (ReA-GRPO) that enhances length extrapolation capability while dynamically determining the number of rereading passes based on task complexity.
Experimental Validation: Extensive experiments demonstrate that MemReread consistently outperforms baseline frameworks on long-context reasoning tasks, while maintaining linear time complexity with respect to context length.

Improvements for AI systems

This system can recover indirect facts that were prematurely discarded by standard memory overwriting by triggering question decomposition and rereading when the final memory is insufficient, as stated in Figure 4 and Section 3.1. This enables the recovery of indirect facts that were prematurely discarded.
The agent can maintain non-linear reasoning capabilities while preserving the inherent logical flow of document comprehension, supporting complex reasoning tasks that require cross-chunk dependencies.
A reinforcement learning framework allows for dynamic control over computational overhead by determining the number of rereading passes based on task complexity, optimizing the performance-efficiency trade-off.
The system can perform targeted rereading guided by generated sub-questions, which facilitates the recovery of discarded facts without inducing interferences, as demonstrated in Figure 1 and Section 2.3.
The framework can implement an adaptive mechanism where Rereading-Adaptive GRPO encourages answering questions with minimal reading passes without compromising accuracy, while reducing the penalty for additional reading on challenging problems.

Abstract

Long-context reasoning tasks have motivated agent-memory approaches that process document chunks sequentially while maintaining a dynamically updated memory, avoiding the quadratic complexity of standard attention. Existing methods often rely on retrieval to recover information discarded during memory overwriting, but remain vulnerable to evidence loss during memory formation and interference from invalid queries. To overcome these limitations, we propose MemReread. Built upon streaming reading, it circumvents intermediate retrieval, triggering question decomposition and targeted rereading when the final memory is insufficient. By recovering indirect evidence that may have been prematurely discarded, MemReread enables non-linear reasoning while preserving the logical flow of document comprehension. We further introduce a reinforcement learning framework that improves length extrapolation and adaptively determines the number of rereading passes according to task complexity, enabling flexible control of computational overhead. Our MemReread-4B outperforms strong baselines on long-context reasoning tasks (e.g., +1.6pp v.s. DeepSeek-V4-Pro on 2WikiMultiHopQA) while maintaining linear time complexity with respect to context length.