Reproduce “ENGRAM: Effective, Lightweight Memory Orchestration for Conversational Agents” on OpenResearch

About this paper

Daivik Patel, Shrenik Patel

ENGRAM, a lightweight memory orchestration system, enables conversational agents to maintain long-horizon consistency by categorizing dialogue history into episodic, semantic, and procedural memory types. This approach utilizes minimal routing and dense retrieval to achieve state-of-the-art semantic accuracy on dialogue benchmarks while significantly reducing context token usage and response latency.

Key insights

• Careful memory typing and minimal routing combined with straightforward dense retrieval are sufficient to achieve state-of-the-art performance in long-term conversational memory tasks, challenging the need for complex architectures.
• Architectural simplicity, when guided by principled design choices for memory organization, can lead to superior performance and practical benefits in managing conversational memory for LLMs.
• Explicit separation of memories into episodic, semantic, and procedural types significantly enhances semantic correctness by concentrating relevant evidence and reducing competition during retrieval, as confirmed by ablation studies.

Summary by alphaXiv.