DecentMem Dual-Pool Memory Halves Token Use and Boosts Collaboration

Problem with Centralized Memory

Multi‑agent systems built on large language models typically use a single shared repository for experience, summaries, and interaction history. While this appears logical for synchronizing agents, it introduces several drawbacks.

Loss of specialization: Agents repeatedly draw from the same pool, causing behavior to converge and eroding clear role distinctions.

Scalability and cost: Retrieval cost grows with pool size; each agent must process more irrelevant context, increasing token consumption and slowing execution.

Privacy and security risks: Shared experiences pose significant risks in enterprise or multi‑organization deployments.

Self‑evolution blockage: Continuous improvement requires learning from collective experience while preserving diversity; a centralized design pushes the system toward a single dominant strategy.

These issues become pronounced when tasks are repeated or workflows become complex, leading to diminishing returns as agents repeat similar patterns without true innovation.

DecentMem Decentralized Dual‑Pool Memory

DecentMem equips each agent with its own private dual‑pool memory, separating exploitation and exploration while still supporting cross‑framework collaboration.

Exploitation Pool (E‑pool)

The E‑pool stores distilled experience from previously successful tasks, capturing rich collaborative details. Each memory entry contains:

Action trajectories (task decomposition, direct responses, tool usage).

Collaboration history—who handled what and how information was passed between agents.

Agent’s own annotations and internal decision rationale.

This pool supports a "local‑walk" mechanism: when selected, an agent retrieves similar past strategies and reuses proven methods as priors for action and collaboration.

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 and experiences 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 solutions no longer apply.

Formally, for each agent m , memory is represented as a pair (E_m, X_m) where E_m holds structured entries and X_m holds transient candidates.

Online Router and Feedback Loop

DecentMem does not fix a static exploitation‑exploration ratio. Instead, an online router dynamically decides, for each sub‑task at time t , which pool to use.

The router computes weights for the two pools.

With probability α it selects the E‑pool; with probability 1‑α it selects the X‑pool.

The chosen pool is queried: the agent either retrieves a past successful case or generates new candidates.

After execution, an external LLM acts as a judge, evaluating the trajectory stage by stage and providing feedback on solution quality, collaboration, and pros/cons.

The system updates the pool weights based on this feedback, biasing future decisions toward the more successful pool.

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

Key Findings

Global reachability: The dual‑pool mechanism guarantees agents can reach any feasible strategy in the local solution space; the E‑pool handles local optimization while the X‑pool transports agents to distant possibilities.

Preserved diversity with coordination: Private memories keep specialization intact, and detailed trajectories retain clear "who did what" signals.

Efficiency gains: Agents only access relevant private memories, reducing context redundancy and cutting token usage.

Accelerated self‑evolution: As tasks accumulate, diversity is better retained, leading to compound performance improvements rather than stagnation.

Empirical Results

Compared with the strongest centralized baseline, accuracy improves up to 23.8%.

Against a no‑memory baseline, the gain reaches 52.5%.

Token consumption drops by as much as 49%.

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

Benefits are larger in loosely coordinated, highly stochastic scenarios where diversity is most valuable.

How to Use DecentMem

Wrap existing agents: Insert DecentMem as middleware; implement each agent’s two pools using vector stores or structured databases, with similarity‑based retrieval in the E‑pool via embeddings.

Define memory fragments: Structure entries to include trajectory, annotations, and collaboration metadata; optionally add provenance information for inter‑agent messages.

Implement the router: A lightweight module samples according to current weights (initially equal) and updates them from feedback.

Add an LLM judge: Use a capable model (or a carefully prompted backbone model) to evaluate each stage, focusing on collaboration quality, solution correctness, and improvement suggestions.

Handle persistence: Persist E‑pool entries across sessions; X‑pool entries can be transient, promoting successful candidates to the E‑pool when appropriate.

Best Practices and Potential Extensions

Continuously monitor router weight changes to understand agent preference evolution and aid debugging.

Combine with existing RAG techniques for pool retrieval.

Apply access control and encryption to private memories in production.

Test mixed static‑dynamic workflows to identify scenarios with maximal benefit.

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

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

Experiment with alternative judging models or multi‑criteria feedback.

Integrate with tool‑calling frameworks to enrich trajectory data.

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

Advantages and Caveats

Lower operational cost: Reduced token usage lowers API or inference expenses, especially for high‑throughput applications.

Improved reliability: Dynamic balancing mitigates failures caused by outdated knowledge or repetitive errors.

Easier debugging and maintenance: Rich metadata in private memories makes it straightforward to trace contributions.

Future‑ready: Supports continual learning without retraining the entire system; agents evolve both individually and collectively.

Compliance‑friendly: Suitable for scenarios involving sensitive data or multi‑organization collaboration.

Additional overhead from the judge and router exists, but it is typically outweighed by token savings; careful design of storage and feedback prompts is required.

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