Loading…
Short-lived task memory, long-lived session memory, and shared organizational memory live in different stores with different policies — and the agent retrieves the right slice automatically.
Dumping everything into the prompt doesn't scale. Real memory means different stores for different time horizons, and a retrieval layer that assembles the right context for each step — without making it the agent's problem.
What's happened in the current run. Ephemeral, cheap, scoped to the request.
Conversation-level state that persists across turns with a user. Bounded, consent-gated, expiring.
Shared knowledge the whole platform can draw on — governed, citable, and updated via deliberate pipelines.
The agent gets the slice it needs, not the entire corpus stuffed into a prompt.
Session memory is governed by explicit retention windows and user consent — not 'keep it all'.
Every recalled fact carries its source and timestamp. Claims are citable.
When memory contradicts live data, the agent is instructed to prefer the live source and flag the drift.