Skip to content
atomic-agent docs

Memory

atomic-agent’s memory is a cognitive model, not a transcript dump. Instead of re-reading your whole chat history, it keeps a tidy, searchable memory and surfaces only the relevant bits when needed.

What it stores

  • Profile facts — durable identity and preferences
  • Notes — episodic memory, searchable with SQLite FTS5 (plus optional embeddings)
  • Links — associations between memories
  • Lessons — distilled takeaways from experience
  • Procedures — reusable how-to knowledge
  • Voting & reflection — usefulness is voted on, and the agent reflects in the background between turns

How it stays fast

Recall is bounded and rendered into a predictable prompt budget, so memory never drowns the context window. The design principle is to keep session state outside the model whenever possible.