AuthorsChengxin Yu, Zhaoxin Fan, Faguo Wu et al.
TL;DR
CoMem uses Private Experience Sedimentation, Collective Wisdom Curation, and Parallel Dual-Stream Retrieval to boost MacNet ALFWorld success from 79.85% to 89.55% (+9.70pp).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CoMem targets memory pollution, where flat shared repositories suffer noise accumulation and behavioral homogenization, eventually driving the collective toward mediocrity.
In conventional MAS, indiscriminate cross-agent writing into a single shared pool causes execution noise dilution and loss of specialization, degrading success rates below even No-memory baselines.
HOW IT WORKS
CoMem centers on Private Experience Sedimentation, Collective Wisdom Curation, and Parallel Dual-Stream Retrieval, all tied together by a unified metadata schema and an empirical promotion gateway.
You can think of CoMem like a team with personal notebooks and a curated handbook: agents first refine notes privately, then only proven strategies are promoted into the shared handbook.
This design lets CoMem maintain diverse, role-specific expertise while still building a high-quality group memory, something a plain context window or flat shared log cannot provide.
DIAGRAM
This diagram shows how CoMem retrieves and combines private and collective memories in parallel when an agent faces a new task.
DIAGRAM
This diagram shows how CoMem is plugged into different MAS frameworks, evaluated on ALFWorld and PDDL, and ablated over key hyper-parameters.
PROCESS
Private Experience Sedimentation
CoMem lets each agent extract its execution trace and use the Reflection Distiller to convert raw trajectories into concise insights stored in Private Experience Sedimentation slots.
Collective Wisdom Curation
CoMem applies Collective Wisdom Curation with the empirical promotion gateway, checking usage count and above-baseline rewards before moving insights into the Collective Memory Pool.
Parallel Dual-Stream Retrieval
During new tasks, CoMem runs Parallel Dual-Stream Retrieval, pulling top k items from both private memory and the collective pool, then applying clustering-based diversity filtering.
Hybrid Utility Tracking and Pruning
CoMem updates hybrid utility scores via EMA, increments miss counters, and performs rolling pruning in private memory while purging low-score entries from the collective pool.
KEY CONTRIBUTIONS
CoMem Architecture for Memory Pollution
CoMem introduces a two-tier group memory with Private Experience Sedimentation and Collective Wisdom Curation, lifting MacNet ALFWorld success from 79.85% to 89.55% (+9.70pp).
Collective-Individual Memory Synergy
CoMem formalizes collective-individual memory synergy by combining isolated private stores, a curated Collective Memory Pool, and Parallel Dual-Stream Retrieval for balanced individual and group learning.
Systematic Evaluation Across MAS Frameworks
CoMem is evaluated on ALFWorld and PDDL with AutoGen, DyLAN, MacNet, and CARD, showing average absolute gains of 8.02 percentage points on ALFWorld and 4.70 on PDDL over No-memory baselines.
RESULTS
ALFWorld MacNet Success Rate
89.55%
+9.70pp over No-memory MacNet (79.85%)
PDDL MacNet Success Rate
70.19%
+9.41pp over No-memory MacNet (60.78%)
ALFWorld AutoGen Success Rate
88.81%
+9.71pp over No-memory AutoGen (79.10%)
PDDL AutoGen Success Rate
74.44%
+4.92pp over No-memory AutoGen (69.52%)
These metrics come from ALFWorld and PDDL, which test embodied control and symbolic planning respectively. The gains show that CoMem’s collective-individual memory synergy reliably improves multi-agent success over strong No-memory baselines.
BENCHMARK
These metrics come from ALFWorld and PDDL, which test embodied control and symbolic planning respectively. The gains show that CoMem’s collective-individual memory synergy reliably improves multi-agent success over strong No-memory baselines.
BENCHMARK
Success rate (%) on ALFWorld for MacNet with different memory mechanisms.
BENCHMARK
Success rate (%) on PDDL for AutoGen with different memory mechanisms.
KEY INSIGHT
On DyLAN with PDDL, G-Memory drops success from 67.78% (No-memory) to 59.53%, while CoMem raises it to 69.50% (+9.97pp over G-Memory).
This is surprising because adding shared memory is usually expected to help, but flat cooperative memory actually harms performance until CoMem’s curated collective-individual design is used.
WHY IT MATTERS
CoMem unlocks stable, long-term group memory where agents keep specialized private knowledge while still benefiting from rigorously curated collective wisdom.
Builders can now deploy LLM-based multi-agent systems that learn across tasks without collapsing into noisy, homogenized behavior, enabling sustained specialization and reliable shared strategies.
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…