AuthorsJingyi Kang, Chunyu Li, Ding Chen et al.
TL;DR
MemReader uses a ReAct-style memory manager with GRPO-optimized tool decisions to reach 83.00% Overall on LongMemEval, matching EverMemOS while using fewer tokens.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing systems treat memory extraction as a one shot passive transcription from context to structured entries, which struggles with noisy dialogue, missing references, and cross turn dependencies.
This causes memory pollution, weak knowledge updating, and poor temporal reasoning, making long term agents unreliable for maintaining user state over time.
HOW IT WORKS
MemReader uses MemReader-4B, MemReader-0.6B, a ReAct-style trajectory, and a memory-operation tool space to turn extraction into explicit memory-state decisions.
You can think of MemReader like a CPU with RAM and disk plus a smart OS scheduler, deciding whether to cache, write, or discard each piece of interaction.
This KEY_MECHANISM lets MemReader maintain a low-noise, updatable memory state that a plain context window or passive JSON writer cannot provide.
DIAGRAM
This diagram shows how MemReader-4B runs a think–act–observe loop over memory tools at each dialogue turn.
DIAGRAM
This diagram shows how MemReader is trained with SFT and GRPO and then evaluated on LOCOMO, LongMemEval, and HaluMem.
PROCESS
Decision State Definition
MemReader constructs the decision state s_t = (x_t, M_{t-1}, B_{t-1}), combining the current utterance with long term memory and buffer.
ReAct Trajectory Generation
MemReader generates a ReAct-style trajectory τ_t with internal reasoning traces and tool actions drawn from the memory-operation tool space.
Tool Action Execution
MemReader executes add_memory, buffer_memory, search_memory, or ignore_memory, receiving observations that inform subsequent reasoning steps.
Memory State Transition
MemReader applies the transition operator T to update long term memory M_t and buffer B_t, maintaining a low-noise, updatable user state.
KEY CONTRIBUTIONS
Active Memory Management with ReAct
MemReader formulates memory extraction as a sequential decision process using ReAct-style trajectories and a specialized memory-operation tool space, achieving 83.00% Overall on LongMemEval.
MemReader-4B with GRPO Optimization
MemReader-4B uses GRPO with multi-level rewards for format, action alignment, content quality, and efficiency to improve knowledge updating, temporal reasoning, and hallucination resistance.
MemReader-0.6B Lightweight Structured Extractor
MemReader-0.6B distills structured memory extraction into a 0.6B Qwen3 model, surpassing a GPT-4o-mini-based baseline on LOCOMO F1 with 52.54 vs 51.90.
RESULTS
Overall (LongMemEval)
83.00%
+0.00 over EverMemOS while using 922 vs 2800 tokens
Knowledge Update (LongMemEval)
91.03%
+1.29 over EverMemOS
Temporal Reasoning (LongMemEval)
84.21%
+6.77 over MemOS
Memory Extraction F1 (HaluMem)
98.21%
+18.51 over MemOS
These numbers come from LongMemEval and HaluMem-Medium, which test long term preference tracking, temporal reasoning, knowledge updates, and hallucination resistance. MAIN_RESULT shows that MemReader can match EverMemOS’s Overall score while cutting average extraction tokens from 2800 to 922.
BENCHMARK
These numbers come from LongMemEval and HaluMem-Medium, which test long term preference tracking, temporal reasoning, knowledge updates, and hallucination resistance. MAIN_RESULT shows that MemReader can match EverMemOS’s Overall score while cutting average extraction tokens from 2800 to 922.
BENCHMARK
Overall score on LongMemEval across memory systems.
BENCHMARK
Memory Extraction F1 on HaluMem-Medium.
KEY INSIGHT
MemReader-4B-GRPO reaches 83.00% Overall on LongMemEval with 922 tokens per extraction, matching EverMemOS which uses 2800 tokens.
This is surprising because larger token budgets are usually assumed necessary for better long term memory, yet MemReader achieves equal accuracy with roughly one third the token cost.
WHY IT MATTERS
MemReader unlocks long term agents that actively judge value, ambiguity, and completeness before writing, buffering, retrieving, or ignoring memories.
Builders can now deploy memory systems that stay low-noise and updatable over many sessions without relying on massive context windows or expensive general purpose APIs for every extraction step.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…