Working versus long-term memory

A practical distinction between state available during an active task and state persisted for possible use in later tasks or sessions.

Working state can include current context, tool results, temporary files, and checkpoints. Long-term state can include preferences, decisions, histories, embeddings, or records with different retention and review rules. Product implementations vary, and a conversation may itself persist even when the model receives only a summary on the next request.

Builder example

Persisting everything increases privacy, staleness, retrieval-noise, and deletion risks; persisting too little loses continuity and recovery. Retention should follow the record's purpose, authority, sensitivity, provenance, update path, and expected future value rather than one small-reviewed-only rule.

Common confusion: Context-window capacity, conversation history, application state, and product memory are related but separate. A long request can be ephemeral, while a tiny preference can persist for years.