Dual Process Memory Architecture for Scientific LLM Agents

A recent study presents a Dual Process Memory Architecture aimed at addressing context window saturation in Large Language Models (LLMs) for extended scientific tasks. This architecture differentiates between short-term episodic requirements, maintained within a fixed 10-message window, and long-term accumulated knowledge, which increases at approximately 3 tokens per message. In contrast to previous memory systems for social agents, this specialized method effectively manages conflicting parameter changes, facilitates multi-hop reasoning throughout various experimental stages, and ensures accurate retention of technical facts. The architecture underwent testing with 15,000 messages and was validated across six LLMs belonging to three different families: OpenAI, Anthropic, and Google.