Persistent memory adoption

AgentPlat 1.4.0 adds an opt-in persistent memory profile with stable-agent subjects and separately authorized personal or shared audiences. Corrections and retirement create immutable revisions. Execution evidence can pin the exact memory and reviewed definition revisions used for a run.

Memory does not infer a private owner from a tenant, participant or conversation ID. The host authenticates a verified actor, maps it to a stable agent, and separately authorizes shared access. Existing legacy memory remains readable through legacy APIs; migration does not automatically convert old entries or infer their audience.

PostgreSQL setup

The PostgreSQL memory ledger is separate and opt-in; it does not advance the Room migration ledger. A memory-only adopter applies the packaged memory revision migration through the explicit migration runner, then checks its status before wiring PostgresMemoryRevisionStoreV1 and MemoryRevisionServiceV1. Adapter construction does not migrate automatically. Follow the exact migration and API instructions in the published consumer guide (opens in a new tab).

Do not backfill old entries without an explicit, reviewed mapping to a stable agent and verified personal/shared audience. Retirement is a revision, not physical erasure. Apply retention and access controls to history, execution snapshots, exports and backups. Rollback is destructive and should not be used as an ordinary correction path.

Definition review and execution evidence

Reviewed definition changes reuse existing Room approvals and governance history; they do not create a new authority owner. Preserve the exact proposal, approval, evaluation and definition revision coordinates. A run can record the effective context digest and exact memory revision references; a digest alone does not prove what a remote provider executed.

The published consumer checks cover isolated PostgreSQL behavior, public types, authorization cases, revision conflicts and reopen/concurrency scenarios. They use deterministic fixtures, not live model behavior, product identity systems or production databases. They establish bounded integration, not production-scale validation.

See also Planner adoption, artifact adoption, and the package catalog.