Agent memory specification
Agent Context Memory
Persistent memory is not enough. An agent also needs to know whether it is authorized to act.
Memory as an authorization problem
Agent Context Memory treats the pre-work mandate as a condition of authorized work, rather than context an agent might consult after starting. Its proposed mandate gate requires principal-authored intent to exist and be read before a valid work session begins. This applies familiar authorization-before-action reasoning to agent memory rather than focusing only on what an agent can recall.
Separate records by trust
The specification distinguishes an intent tier, a trace tier and an independently captured evidence tier. Intent belongs to the principal. Trace holds the agent’s decisions and orientation. Evidence comes from a capture mechanism outside the agent’s authorship. Keeping those roles distinct prevents an elegant retrospective explanation from silently becoming the authority that supposedly permitted the action in the first place.
Scope and continuity
The familiar .acm directory pattern places a destination, an append-only audit trail and a derived orientation beside the target being worked on. The specification also discusses nested scopes so that higher-level mandates govern narrower ones. This supports continuity across sessions while keeping durable direction distinct from a particular agent’s current interpretation of it.
Specification versus enforcement
Files and naming conventions alone cannot establish that an implementation prevents unauthorized writes or protects captured evidence. Those properties depend on the actual access controls, capture mechanism and host integration. ACM provides a specification and reference pattern to examine against those requirements; this overview does not certify every .acm directory as compliant. The public specification and citation metadata are the authoritative starting point for its detailed rules and relationship to AII and PEA.