FSFM: A Biologically-Inspired Framework for Selective Forgetting of Agent Memory

Selective forgetting is as important as memory retention for AI agents: The paper argues that in resource-constrained environments, intelligently designed forgetting mechanisms are equally critical to what AI systems remember, challenging the field's traditional focus on memory retention alone.

Forgetting improves efficiency, quality, and security simultaneously: Empirical results demonstrate that selective forgetting delivers measurable improvements across three key dimensions—access efficiency (+8.49%), content quality (+29.2% signal-to-noise ratio improvement), and security performance (100% elimination of security risks in tested scenarios).

Four types of forgetting mechanisms serve different purposes: The framework identifies passive decay-based forgetting, active deletion-based forgetting, safety-triggered forgetting (for removing malicious or sensitive data), and adaptive reinforcement-based forgetting, each addressing different responsible AI concerns.

Selective forgetting addresses privacy and security risks directly: The paper highlights that active forgetting of malicious inputs, sensitive data, and privacy-compromising content is a practical mechanism for compliance with regulations and protection against misuse—moving beyond traditional security approaches.

The approach is inspired by human cognition but adapted for AI systems: Drawing from neuroscience concepts like hippocampal consolidation and the Ebbinghaus forgetting curve, the framework translates biological memory principles into concrete technical implementations suitable for modern AI agent architectures.