Biologically-Inspired Selective Forgetting Framework for LLM Agents

A novel framework named FSFM (Forgetting for Selective Memory) introduces a biologically-inspired approach to selective forgetting for LLM agents, utilizing concepts from hippocampal indexing/consolidation theory and the Ebbinghaus forgetting curve. This framework posits that in environments with limited resources, the act of forgetting is as vital as remembering, enhancing efficiency (through memory pruning), quality (by updating outdated preferences), and security (by eliminating malicious inputs and sensitive information). It categorizes forgetting mechanisms into four types: passive decay-based, active deletion-based, safety-triggered, and adaptive reinforcement-based. This research builds on recent developments in LLM agent architectures and vector databases, with comprehensive specifications provided. The paper can be accessed on arXiv with ID 2604.20300.