Skip to content

Chapter 8

Remembering and Forgetting: How Agents Manage Experience

Interactive Graph (beta)

Toggle graph

Inspired by Human Memory Systems in Cognitive Neuroscience

Human intelligence relies on multiple memory systems and strategic forgetting. Agents need the same: memory bridges moments, connects experiences, and enables adaptation. Without it, an agent is a goldfish; with it, an agent is coherent and learning.

Human Analogy: Different Kinds of Memory

  • Working Memory (short-term focus): like an LLM’s context window.
  • Episodic Memory (personal history): past conversations, tasks, actions.
  • Semantic Memory (facts/knowledge): knowledge bases or vector stores.
  • Procedural Memory (skills/habits): rules, prompts, refined behaviors.
  • Forgetting (selective pruning): summarizing or discarding to stay efficient.

Why Agents Need Memory

Memory supports long conversations, multi-step progress tracking, personalization, and learning from experience.

Everyday Parallel

Like a teacher who knows each student’s history, an agent tutor adapts based on prior mistakes and concepts struggled with.

Design Lessons for Agents

  1. Balance immediacy and persistence.
  2. Structure memory into layers.
  3. Forgetting is essential for efficiency.
  4. Memory enables identity and continuity.

Simple Flow

user_input -> short_term_memory
if relevant: promote to long_term_memory
if outdated: summarize or discard
long_term_memory -> retrieved into short_term when needed

Conclusion

Agents thrive by remembering enough to be coherent and adaptive, and forgetting enough to stay efficient.