Memory that outlives the runtime.
Continuity that stays with the agent.

Engram runs a heterogeneous model stack — not one model for everything, but the right model for each cognitive task.

Triage — fast, cheap models decide in milliseconds what's worth remembering. Most inputs are skipped here.

Conversation — reasoning-capable models handle the actual agent response with full memory context injected.

Calibration — powerful models extract structured facts from conversations and write them back to Engram.