DecentMem’s Dual‑Pool Memory Cuts Token Usage by Almost 50%

Problem with Centralized Memory in Multi‑Agent LLM Systems

In multi‑agent systems built on large language models, all agents typically read from and write to a single shared repository. While this appears logical for synchronizing information, it creates four major issues:

Loss of agent specialization : agents repeatedly draw from the same memories, causing behavior to converge and eroding distinct roles.

Scalability and cost : retrieval cost grows with the size of the shared pool, increasing token consumption and slowing execution.

Privacy and security risks : shared experience can expose sensitive data in enterprise or multi‑organization deployments.

Self‑evolution blockage : continual reliance on collective experience pushes the system toward a single dominant strategy, hindering long‑term improvement.

These problems become especially visible when tasks are repeated or workflows become complex, leading to diminishing returns.

DecentMem’s Decentralized Dual‑Pool Memory

Exploitation Pool (E‑pool)

The E‑pool stores distilled experiences from successfully completed tasks. Each memory entry contains:

Action trajectory (task decomposition, direct response, tool usage).

Collaboration history – which agent performed what and how information was passed.

Agent‑specific annotations and internal decision rationale.

A "local‑walk" mechanism retrieves similar past strategies, allowing agents to reuse proven methods as a prior for current actions and coordination.

Exploration Pool (X‑pool)

The X‑pool acts as a temporary buffer for the current task. It invokes the LLM’s generative capability to create new candidate solutions for unseen situations, avoiding reliance solely on history.

This "heuristic teleportation" jumps to new regions of the solution space, preventing agents from getting stuck in local optima when old strategies no longer apply.

Online Router and Feedback Loop

DecentMem does not fix a static exploitation‑exploration ratio. An online router dynamically decides, for each sub‑task, whether to draw from the E‑pool or the X‑pool:

The router computes weights for the two pools.

With probability α it selects the E‑pool; otherwise it selects the X‑pool.

The chosen pool provides either a retrieved past case or a newly generated candidate.

An external LLM acting as a judge evaluates the trajectory stage‑by‑stage, giving feedback on solution quality, collaboration, and shortcomings.

The system updates the pool weights based on the feedback, biasing future decisions.

This bandit‑style online learning yields an O(log T) cumulative regret bound, matching the theoretical lower bound for stochastic bandits.

Key Empirical Findings

Against the strongest centralized baseline, accuracy improves by up to 23.8 % .

Compared with a memory‑less approach, the gain reaches 52.5 % .

Token consumption drops by as much as 49 % .

Evolution speed in frameworks such as DyLAN is roughly 2.5 × faster.

Benefits are larger in loosely coordinated or highly stochastic collaboration scenarios, where diversity matters most.

Tests cover AutoGen, DyLAN, and AgentNet frameworks, multiple model families (Qwen‑3 4B/8B/14B, Gemma‑4), and a range of benchmarks (math, code, QA, embodied tasks).

How to Apply DecentMem

Integration Steps

Wrap existing agents with DecentMem middleware; implement the two private pools using vector stores or structured databases. Use embeddings for similarity search in the E‑pool.

Define memory fragments to include trajectory, annotations, and provenance metadata.

Implement a lightweight router that samples according to the current weights (initially equal) and updates them from feedback.

Add an LLM‑as‑judge module that evaluates each stage, focusing on collaboration quality, solution correctness, and improvement suggestions.

Persist E‑pool entries across sessions; optionally promote successful X‑pool entries to the E‑pool.

Best Practices

Continuously monitor router weight changes to understand evolving agent preferences.

Combine with existing Retrieval‑Augmented Generation (RAG) techniques for efficient intra‑pool search.

Apply access control and encryption to private memories in production.

Test both fixed‑ratio and dynamic‑ratio workflows to identify the most beneficial scenario.

Distribute agents across multiple nodes; decentralized memory naturally fits distributed deployments.

Potential Extensions

Introduce decay or summarization mechanisms for the E‑pool to improve long‑term efficiency.

Swap in alternative judging models or multi‑objective feedback signals.

Integrate with tool‑calling frameworks to enrich trajectory data.

When privacy permits, enable selective cross‑agent sharing of high‑value distilled memories.

Advantages for Development

Lower operating cost : fewer tokens mean reduced API or inference expenses, especially for high‑throughput applications.

Higher reliability : dynamic balancing reduces failures caused by outdated knowledge or repetitive errors.

Easier debugging and maintenance : private memories carry rich metadata, making it straightforward to trace which agent contributed which information.

Future‑proof self‑learning : agents can evolve individually and collectively without retraining the entire system.

Compliance‑ready : isolated memories suit scenarios with sensitive data or multi‑organization collaboration.

Conclusion

Multi‑agent LLM systems are moving toward genuine self‑evolution, and memory architecture is the decisive factor. DecentMem demonstrates that decentralizing memory while preserving structure and adaptability resolves the tension between collaboration, specialization, and efficiency. By combining private dual pools with an online router, the framework achieves substantial accuracy gains, token savings, and faster learning without sacrificing coordination.

Paper: https://www.alphaxiv.org/abs/2605.22721

Code example

来源：DeepHub IMBA
本文
约3000字
，建议阅读
5
分钟
DecentMem 证明：把记忆去中心化，同时保持结构化和自适应能力，能够化解协作、专业化与效率之间的张力。