State with boundaries
Agents Cache Index
Keep reusable observations separate from decisions, memory, and actions.
The idea, in context
An agent may read documents, call tools, remember context, and act on results. An agents cache index describes reusable observations within that process. It should not blur a past tool result with current authority to act. A retained stock count, permission check, or workflow state may be out of date even if its text is unchanged. Treat the action boundary as a separate checkpoint and decide which observations require fresh validation before the agent proceeds.
Classify each tool result
Describe whether the result is a stable reference, a time-sensitive observation, or evidence tied to one action. Reuse of a published manual can follow different rules from reuse of a reservation status. Put the classification in the tool contract.
Version the surrounding state
Include the tool definition, relevant parameters, source version, and permission scope in eligibility checks. If the agent’s instructions change, review which retained results still make sense. A namespace can separate one workflow revision from another.
Recheck before consequential steps
A cached observation can help plan a task, but planning and execution can be separated in time. Define the fresh checks required before an action. Record why the agent reused or refreshed a result so later review can follow the decision.
Questions about agents cache index
Is agent memory just a long-lived cache?
It can share storage techniques, but the purpose differs. Memory carries context across interactions; a cache avoids repeated work under defined conditions. Design deletion, retrieval, and reuse policy for each purpose.
Can an agent cache permission checks?
Treat permission information as especially time-sensitive. An earlier successful check should not be the sole basis for a later action after access may have changed.
Start with one good question.
Follow a topic, explore the tradeoffs, and go deeper in the Lab.