New Framework Enables LLM Agents to Share and Learn from Collective Memory

Large language model-based agents can handle intricate tasks, yet their effectiveness in unpredictable environments often suffers from working in isolation and drawing from fixed, unchanging knowledge bases. Unlike how humans naturally exchange insights through conversation, these AI systems typically lack dynamic ways to share information.

The Challenge#

Traditional multi-agent setups rely on static databases and don't capture how information naturally flows between participants. This limitation prevents the kind of continuous learning and knowledge refinement that happens in real human collaboration.

INMS: Interactive Memory Sharing#

A research team has introduced the INteractive Memory Sharing (INMS) framework to address this gap. Rather than agents operating independently, INMS creates an asynchronous interaction model where multiple agents can tap into a shared conversational memory pool.

The system handles three core functions: filtering incoming information in real time, storing it systematically, and retrieving it when needed. By doing so, INMS enables agents to engage in ongoing dialogue-like exchanges that mirror how human teams build collective knowledge.

How It Strengthens Performance#

Testing across three different datasets showed that the framework improves agent capabilities by better modeling how multiple agents interact and share information. The system continuously refines its retrieval mechanisms based on past interactions, allowing it to grow smarter as agents collaborate.

The result is collective self-enhancement—agents don't just solve individual tasks more effectively, but the group as a whole learns and improves from shared experience.