A Memory Event is the primary worker-emission object. It is a proposal envelope around observable task evidence. A Memory Ticket can be a human-facing queue wrapper, but the canonical schema is the Memory Event.
Core Scope Model
| Scope | Owner | Allowed contents | Retrieval rule | Leakage failure |
|---|---|---|---|---|
| session | Run owner | Temporary observations, hypotheses, incomplete task state. | Excluded from broad durable retrieval unless promoted. | Ephemeral guess retrieved as project fact. |
| project | Project steward | Project decisions, risks, procedures, preferences, evidence, current state. | Retrieve only inside the same project/tenant boundary. | Project record leaks to unrelated project. |
| agent_team | Team steward | Shared procedures, handoff standards, routing conventions, safety rules. | Retrieve for the named team after stronger review. | Team preference treated as universal policy. |
| agent_repo | Specialist-agent owner | Specialist procedures, interface contracts, durable role knowledge. | Retrieve for the role or agent identity it belongs to. | Specialist rule applied to the wrong agent. |
| discard | Curator/audit owner | Malformed, unsafe, unsupported, redundant, unauthorized, or out-of-scope events. | Not searchable durable memory; keep auditable reason where policy permits. | Discarded unsafe event is indexed as usable memory. |
Optional Extension Scopes
x_user_identity, x_organization, and x_cross_organization are extension scopes, not core scopes. User-identity memory requires explicit owner, consent or lawful basis where relevant, access rules, deletion/erasure process, retention, source quality, review, and cross-scope retrieval restrictions.
Nine Core Kinds
| Kind | Definition | Evidence and retrieval rule |
|---|---|---|
| fact | Claim about observable or source-backed state. | Needs evidence reference; retrieve as fact only when reviewed/current. |
| decision | Choice with context, authority, rationale, effective scope, and date. | Needs decision owner; retrieve by effective scope and supersession state. |
| preference | Scoped preference, not universal policy. | Needs owner and scope; do not promote across scope silently. |
| risk | Potential failure, harm, uncertainty, or exposure. | Needs impact/trigger; retrieve near related tasks and reviews. |
| procedure | Repeatable workflow with trigger, preconditions, steps, validation, stop conditions, and rollback/escalation. | Needs test or review; retrieve with current/deprecated state visible. |
| hypothesis | Unproven explanation or candidate relationship. | Must not retrieve as fact; expires or requires promotion review. |
| evidence | Source, artifact, result, or observation supporting a claim. | Supports a claim but is not automatically the claim. |
| deprecation | Record that a prior memory is superseded, retired, or unsafe for current use. | Must reference superseded record; retrieval should redirect or warn. |
| conflict | Credible incompatible claims preserved until resolution. | Retrieve both sides with conflict status; do not fake consensus. |
Memory Event Fields
The starter bundle includes JSON Schema draft 2020-12 files for Memory Events, durable records, curation reports, transactive directories, retrieval requests, and retrieval responses. Core Memory Events require schema version, event id, UTC timestamp, source agent, task, artifact, memory kind, suggested scope, proposal-only worker write mode, evidence references, provenance, sensitivity, confidence, and review hints. Hidden reasoning fields and direct durable write states fail validation.
Example Event
{
"schema_version": "llmwikis.matm.memory-event.v1",
"event_id": "evt_demo_001",
"emitted_at": "2026-06-23T17:00:00Z",
"worker_write_mode": "proposal_only",
"durable_write_allowed": false,
"memory_kind": "procedure",
"suggested_scope": "agent_team",
"disposition_hint": "append",
"artifact": {
"artifact_type": "action_observation_trajectory",
"summary": "The retry step succeeded only after the cache key was cleared.",
"hidden_reasoning_included": false
}
}