AuthorsBowen Ye, Yongchao Xu, Zhijian Li et al.
TL;DR
CoEvo-Mem co-evolves a Prior-Guided Residual Router with a Typed Relational Memory Graph, reaching 68.33% on GPQA Diamond (+7.50 over MemRL and MemQ).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CoEvo-Mem targets the issue that existing methods "typically optimize either memory access ... or memory evolution" while ignoring their mutual feedback loop.
When retrieval and memory updates are optimized separately, long-term LLM agents misalign stored utilities and routing, degrading long-horizon reasoning and cross-session performance.
HOW IT WORKS
CoEvo-Mem centers on Self-Routed Query Rewriter, Prior-Guided Residual Router, Q-Weighted Hybrid Retriever, and a Typed Relational Memory Graph updated by Trajectory-Conditioned Memory Update.
You can think of CoEvo-Mem like a CPU with a smart cache: the router chooses which memory channels to read, while the graph reorganizes disk blocks based on usage.
This tight loop lets CoEvo-Mem adapt routing and memory utilities over time, enabling behaviors a fixed context window or static RAG pipeline cannot support.
DIAGRAM
This diagram shows how CoEvo-Mem processes a single query through rewriting, routing, hybrid retrieval, and answer generation.
DIAGRAM
This diagram shows how CoEvo-Mem alternates Router and Memory phases over the training set D_tr.
PROCESS
Self-Routed Query Rewriter
CoEvo-Mem uses Self-Routed Query Rewriter to generate dense and sparse rewrites plus a routing prior that conditions the Prior-Guided Residual Router.
Prior-Guided Residual Router
CoEvo-Mem feeds rewrites into the Prior-Guided Residual Router, which samples or outputs dense–sparse mixture weights guided by task rewards.
Q-Weighted Hybrid Retriever
CoEvo-Mem applies the Q-Weighted Hybrid Retriever over the Typed Relational Memory Graph, fusing dense, sparse, and utility ranks into the exposure set A_t.
Trajectory-Conditioned Memory Update
CoEvo-Mem runs Trajectory-Conditioned Memory Update to assign TD-style utility residuals and propagate credit across the Typed Relational Memory Graph.
KEY CONTRIBUTIONS
Coupled Retrieval–Memory Learning
CoEvo-Mem formalizes retrieval routing and memory evolution as a coupled learning problem, with the Prior-Guided Residual Router and Typed Relational Memory Graph co-adapted via alternating phases.
Prior-Guided Residual Routing
CoEvo-Mem introduces Self-Routed Query Rewriter plus Prior-Guided Residual Router, learning only a lightweight residual over the LLM prior using Beta-policy reinforcement learning.
Retrieval-Aligned Relational Memory Evolution
CoEvo-Mem designs Trajectory-Conditioned Memory Update on a Typed Relational Memory Graph, propagating TD-style utility across Dense, Sparse, and Time edges with up to 10.00 point gains in ablations.
RESULTS
LLAB success rate
76.00%
+1.56 over MemQ
LiveCodeBench score
55.24%
+3.81 over MemQ
MMMU Pro accuracy
55.49%
+1.06 over MemQ
GPQA Diamond accuracy
68.33%
+7.50 over MemRL and MemQ
Across LLAB, LiveCodeBench, MMMU Pro, and GPQA Diamond, CoEvo-Mem consistently leads the benchmark table, showing that co-evolving routing and memory utilities improves diverse long-term agent tasks.
BENCHMARK
Across LLAB, LiveCodeBench, MMMU Pro, and GPQA Diamond, CoEvo-Mem consistently leads the benchmark table, showing that co-evolving routing and memory utilities improves diverse long-term agent tasks.
BENCHMARK
Mean accuracy or success rate (%) for CoEvo-Mem and baselines on LLAB.
BENCHMARK
Accuracy (%) of CoEvo-Mem variants on GPQA Diamond.
KEY INSIGHT
On GPQA Diamond, CoEvo-Mem reaches 68.33%, while MemRL and MemQ both plateau at 60.83%, a +7.50 percentage point gap.
This is surprising because MemRL and MemQ already learn memory utilities from reward, yet CoEvo-Mem still gains by jointly adapting routing and relational credit propagation.
WHY IT MATTERS
CoEvo-Mem shows that retrieval policies and memory banks should be trained as a coupled system, not as independent modules.
Builders can now design long-term agents where query routing, memory valuation, and graph structure co-evolve, enabling robust performance across heterogeneous tasks without fine-tuning the base LLM.
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.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…