Wiki / concepts / wiki

Agent memory types

The Cognitive Architectures for Language Agents (CoALA) paper sorts agent memory into four kinds. Working memory is the context window the agent reasons over. Episodic memory is the raw record of past behaviour, such as trajectories, tool calls and prior conversations, while semantic memory holds facts about the user and the world. Procedural memory is skills, the knowledge of how to complete a task.

michael-ryan's examples (Stanford CS329Z, Oct 2026): the Generative Agents memory stream for episodic memory, ChatGPT's stored facts about a user for semantic memory, and Anthropic's brand-guidelines skill file for procedural memory (agent-skill-file). Ryan argues for keeping raw episodic logs even though the other types are distilled from them, because the distillation is lossy and the logs let you search again if you chose badly what to remember. Ryan also says the boundaries are fuzzy.

Each type can live in the weights or be retrieved into the context (params-vs-context-memory). Storage options are a file system searched with bash, a database (vector, SQLite, BM25) or a knowledge graph such as Zep, which mostly holds semantic memory.

Source: report

Linked from

Michael RyanParameters vs context memory