MemCog: Memory-as-Cognition System for Conversational Agents
A team of researchers has introduced MemCog, an innovative memory system designed for conversational agents that integrates memory access into the reasoning process, moving away from the conventional Memory-as-Tool approach. MemCog structures user information within a Navigable Memory Store, utilizing associative link graphs, and facilitates multi-step reasoning through a Cross-Dimensional Navigation Interface. Additionally, it employs a Proactive Reasoning Protocol to encourage spontaneous memory retrieval based on conversational context. They also present ProactiveMemBench, the inaugural benchmark for assessing proactive memory activation. Experimental results indicate that MemCog excels on passive QA benchmarks, achieving scores of 92.98 on LoCoMo and 95 on another unnamed benchmark. The research paper can be found on arXiv with ID 2605.28046.
Key facts
- MemCog proposes a Memory-as-Cognition paradigm.
- Existing systems follow a Memory-as-Tool paradigm with one-shot retrieval.
- MemCog includes a Navigable Memory Store with associative link graphs.
- It uses a Cross-Dimensional Navigation Interface for multi-step traversal.
- Proactive Reasoning Protocol drives spontaneous memory exploration.
- ProactiveMemBench is the first benchmark for proactive memory triggering.
- MemCog achieved 92.98 on LoCoMo benchmark.
- The paper is on arXiv with ID 2605.28046.
Entities
Institutions
- arXiv