Agent Memory

Agent Skills

Search and use procedural memory created from successful Agent behavior.

Agent Skills are versioned procedures that describe how an Agent can complete a task. Search Skills with terms from the task, then read a matching Skill before you apply it.

Search Skills

Skill search uses OR semantics for lexical terms. A Skill matches when it contains any search term. Exact package name, Skill ID, and Skill name matches rank before lexical matches.

Use a task family to keep results within the relevant type of work:

results = client.agent.skill.search(
    agent_uuid=agent_uuid,
    query="reservation cancellation",
    task_family="tool.airline",
)

Why did compilation produce no Skill?

List compilation outcomes first:

GET /api/v4/agents/{agent_uuid}/skill-compilations/list

Check the stage and decision in each result.

  • no_eligible_trajectories means that no Trajectory met the compilation requirements.
  • reflector / no_change with no rejection class means that the selected evidence did not produce a proposal.
  • reflector_invalid_output means that the reflector returned invalid output.
  • applicability_ungrounded means that a proposed applicability condition was not supported by the Trajectory evidence.
  • corroboration_insufficient means that the evidence did not support the proposal with enough independent observations.
  • scanner_rejected means that an artifact failed scanner validation.
  • curator_rejected means that the curator rejected the proposal.
  • curator_attempts_exhausted means that the curator could not produce an accepted result before the attempt limit.
  • prompt_budget_exceeded means that a rendered compiler prompt exceeded the gateway byte budget.
  • provider_error means that an external model or provider request failed.

Every outcome includes a compilation_identity. The list is ordered by descending numeric ID. A repeated compilation write with the same identity returns the existing outcome.

If the outcome is admission / produced and no promoted Skill exists, the compiler produced candidates for review. Use the candidate review endpoints to inspect those candidates.