Agent Memory

A deployed model does not normally update its weights after every conversation. The application supplies history or retrieves stored records on the next call. The store can be a database, a file, a graph, or another persistence mechanism.

Our memory explainer distinguishes working context, semantic facts, episodic experiences, and procedural guidance. These are useful architectural categories, rather than a mandatory design for every agent.

Example: a support agent may retain a customer’s preference as a fact, the last unresolved interaction as an episode, and an escalation workflow as procedural guidance. Each needs an owner, a source, access rules, and a way to handle corrections or expiry. Remembering an outdated preference can be worse than starting fresh.

Keeping records between sessions does not make them usable in another application. Agent memory portability adds transfer, access and update checks; the portability guide walks through a small test.

Richmond Alake explains his approach in episode 53, and Conor’s “Your AI Agent has an Amnesia Problem” connects memory with context and knowledge infrastructure. Retrieval ranking can reduce the visibility of old records, while retention and deletion remain separate product decisions.