Agent-Managed Memory Systems

Dated snapshot — verified 2026-08-27

Representative system: Letta, carrying the MemGPT lineage.

The architecture

An operating-system analogy: the context window is RAM, an external store is disk, and the agent itself manages the boundary through tool calls. The model decides what to page in, what to write out, and what to edit — memory operations are actions the agent takes, not a pipeline that runs around it.

What it maps to in this course

The promotion decision, handed to the model instead of a gate. Also the strongest contrast with the system-managed approach every other architecture here takes.

What to look at critically

The trade is control for autonomy. Agent-managed memory adapts to situations a fixed pipeline handles badly, and it is correspondingly harder to audit, test, and constrain — the same properties that make a promotion gate boring make it verifiable. Ask how you would write a regression test for a memory decision the model makes differently each run, and whether the system records why it wrote what it wrote.