Learning outcomes
- Distinguish state from context
- Choose memory lifetimes
- Prevent stale memory
Mental model
State is durable application data. Memory is a policy for creating, retrieving, updating, expiring, and presenting selected state to a model.
Theory
Working state supports the current task; episodic records capture prior interactions; semantic memory stores extracted facts; procedural memory represents stable instructions or workflows. Store authoritative facts in typed systems, not only transcripts. Each memory needs provenance, scope, consent, freshness, and deletion behavior.
Alternatives and trade-offs
Conversation replay is simple but noisy. Summaries compress history but can drift. Structured state is precise. Retrieval-based memory scales but requires relevance and deletion controls.
Failure modes and misconceptions
Do not equate a chat log with memory, persist sensitive inferences silently, accept model-written facts without validation, or allow stale memory to override current input.
Knowledge check
Why is memory a policy rather than merely a database table?
Decision scenario
A learning app stores explicit bookmarks and private notes locally, while inferred preferences remain temporary unless the learner chooses to retain them.
Relationships
Primary sources
- Effective Context Engineering for AI Agents - Anthropic, verified 2026-07-16