Memory

Memory is what an agent retains across sessions: durable scoped memories created by agents and admins. The memory surface is /ai/memory. Each row carries memory, scope, type, agent, status, and update date. Scope is per-agent (Agent: X) or shared (All agents); types are lesson, fact, and preference; rows are active or archived. The header counts active memories, and index errors are a health signal, not background noise.

State each session’s result in one sentence. If you cannot, the session was too broad to remember usefully.
Name sessions so future work finds them: goal plus date beats “session 12”.
Records belong to their tenant and org scope. Never treat work from one tenant as portable to another.
Denials are outcomes, not failures; a denied action that should not have been proposed is a learning signal, so report it.
Until memory surfaces are sourced: assume nothing carries over between sessions except what the record shows. Re-brief the agent per session; do not rely on recall.

Memory: durable scoped memories with per-agent or shared scope, type, and active or archived status

Treat /ai/memory as the source of truth for memory surfaces. Session discipline: Working with sessions. Reporting over the record: Observability.