Capability records
Typed records can represent observations, corrections, decisions, procedures, skills, policies, run results and evidence.
A governance-oriented standard for retaining, evaluating, revising and reusing durable records that allow agent systems to form reusable knowledge and abilities over time.
The standard
ACFS defines the lifecycle and governance model for agent-contributed state. Implementations remain responsible for storage, execution and transport.
Defined capabilities
Typed records can represent observations, corrections, decisions, procedures, skills, policies, run results and evidence.
Records carry identifiers, lifecycle state, scope, authority, provenance and relationship metadata.
Consumers should not apply a record outside its declared scope without authorized broadening.
Submitted, Validated, Rejected, Active, Superseded, Deprecated and Revoked states make behavior governable.
Retrieval can filter by type, scope, lifecycle state, authority, provenance, recency and relation.
ACFS can be transported through MCP, while lifecycle and authority rules remain binding regardless of transport.
Capability model
A single implementation may combine roles, but the standard keeps their responsibilities conceptually distinct.
Governance & security
Implementations should preserve where a record came from and what evidence supports it.
Revoked records must not guide future behavior, while remaining available for audit when appropriate.
Skills, policies, triggers and action-capable records require stronger review than ordinary observations.
The standard calls out unauthorized writes, unsafe promotion, stale records, sensitive persistence and cross-agent propagation risks.
Evidence & status
The repository defines requirement language, capability records, lifecycle states, retrieval rules, MCP relationship and security/trust considerations. It does not claim to be a storage engine or complete agent framework.
The public repository contains the specification and implementation guidance.