Trust & safety
Experience & Chips
An agent that's been used for a while knows more than the day it was built. That earned knowledge is stored separately from the agent itself, as a portable asset called an Experience pack.
What's actually in a pack
Only deltas: a promoted procedure, a verified fix for a failure that kept recurring, a tool gotcha the agent ran into. Never a copy of the agent's own base material — a pack is bound to a specific agent and specific compatible releases, not swapped freely between agents.
Candidate, then promoted — never automatic
- New learnings start as candidates. Nothing is promoted to durable, reusable knowledge on confidence alone.
- A verifier (the Memory Curator, or another declared one) promotes a candidate only after reviewing real evidence for it.
- Retrieval is bounded — at most 8 items, capped around 800 tokens per call. The whole pack is never dumped into a prompt at once.
"Chip" is what you see in the Agent library
In Desktop, a bound Experience pack shows up on an agent as a Chip — separated into what's active in your current conversation versus what applies to newly started ones. It's the visible, ownable form of the same underlying pack.
View your agents' Experience Chips →