Organizational Memory for AI Agents: Department Scopes, Onboarding and Offboarding in a Multi-Tenant Memory Server

Inside a company, the long-term memory of AI agents belongs to departments and outlives the people who wrote it. It is organizational memory, and it needs department boundaries enforced by the server, roles, authorship, versions, and a lifecycle for people who join, move and leave. We describe the team server of total-agent-memory 14.5.1. It gives every personal, team and shared scope its own worker process and SQLite store, and it checks membership on every request. On a synthetic company with four departments and 12 users, 2,532 attack calls across nine attempt types returned 0 foreign records, while a control query found the correct record in the top 5 for 99.5% of 1,200 questions. The same attacks on one store with department tags and no client filter returned foreign-tagged records in 1,176 of 1,176 calls. Membership changes applied from the next request: 0 of 170 requests saw old access. Concurrent edits with 2 to 16 clients lost 0 updates in 400 rounds. The benchmark also found two bugs in how the server handles people who leave. One of them told clients that 16 of 40 saved writes had failed. Both are fixed, with regression tests, in release 14.6.0. Isolation has a cost: with 8 scopes, warm search p50 was 2,938 ms against 459 ms for one store, because every scope ran its own reranker. Release 14.6.0 reranks once in the gateway instead. This cut the 8-scope p50 to 652 ms, within 31% of one store, and left single-scope results unchanged for 400 of 400 questions.