05 · Getting Models to Act

Agent Memory

How an agent remembers across steps

External state a product saves and later puts back into model context: summaries, project files, profiles, database records, or retrieved notes. This is a harness feature, not permanent memory inside the base model.

Concrete example

A research agent writes findings to memory as it goes, then pulls them back when assembling the final report — instead of re-reading every source.

Why it matters

A practical bottleneck for agents that run long, or that need to remember you between sessions.