Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Learning What to Remember: A Cognitively Grounded Multi-Factor Value Model for Agentic Memory
TL;DR
Long-running large language model (LLM) agents face challenges in managing extensive interaction histories, particularly in deciding what to remember or forget. A multi-factor memory value function was developed, incorporating seven cognitive psychology factors to optimize memory management.
✦ Why It Matters
Engineers can implement this multi-factor memory model to improve memory efficiency in AI applications.
Key Takeaways
- Explore multi-factor models for memory management in AI.
- Utilize cognitive psychology factors to enhance memory retention.
- Implement gradient-free optimization for weight learning.
- Evaluate memory strategies using benchmarks like LongMemEval.
Full Summary
As LLM agents accumulate vast interaction histories, they must make critical decisions about what information to retain or discard, which is challenging given the limitations of memory. To address this, a multi-factor memory value function, V(m), was created, utilizing seven factors from cognitive psychology: emotional intensity, goal relevance, value alignment, self/user relevance, task utility, reliability, and usage history.
Weights for these factors are learned through a gradient-free optimization process, allowing for a nuanced approach to memory encoding and retrieval. Experiments on the LongMemEval benchmark demonstrated that this model retained 0.770 of gold evidence, compared to 0.657 for uniform weights and 0.368 for recency-based methods.
Notably, the model's learned weights highlighted the importance of reliability and emotional intensity while appropriately down-weighting goal similarity during forgetting decisions. This research confirms that a tailored multi-factor approach can significantly enhance memory management in AI systems.