R1 · how it works · memory

How R1 remembers.

R1 memory is not one giant chat transcript. It is a tiered store with different retention rules and different audit value.

The homepage memory demo shows the session graph as a live scene. The important operating model is simpler: stable run facts load first, task facts stay close, topical recall is pulled on demand, and durable patterns are written back only after a gate passed.

Tier 0 and Tier 1

Identity and critical facts are what keep the run coherent. Operator stance, task identity, budget rules, security posture, and the current acceptance boundary all belong here. These are the facts R1 treats as load-bearing.

Tier 2

Topical recall is where a task keeps the symbols, prior diffs, and nearby lessons that help with the current phase. This is the tier that changes most while the run is active.

Tier 3

Durable semantic memory is where R1 keeps patterns worth reusing later: regression lessons, known-safe scaffolds, and verified examples. The point is not unlimited recall. The point is selective recall with a proof chain.

See the live memory demo on the homepage.