Context is more than a conversation
Documents, temporary run context, session summaries, personal memory, agent memory, workspace decisions, and organizational knowledge are distinct layers. They carry different owners, permissions, retention rules, and provenance. They are not silently folded into model weights.
A reviewed path to shared knowledge
Agents draft knowledge; an authorized owner reviews publication. Org Brain stores versioned records, source references, and relationships locally. A second authorized agent can retrieve approved context with its provenance, while an unauthorized identity cannot. This is a working local workflow; the customer-owned multi-user server is a later phase.
Boundaries survive reuse
Context requests use authenticated identity and scope checks before retrieval and token budgeting. Restrictions follow source-derived summaries. Revocation, deletion, expiry, and source changes invalidate affected future retrieval and caches. Generated content remains untrusted input to tool execution.
The present agent boundary
Registered agents currently run bounded, owner-started context search and optional task-memory drafts with cancellation and traces. The general inference-driven tool loop and full connector resynchronization are unfinished. Selected read-only Git-file import exists; automatic repository synchronization does not.
Deletion has limits
Deleting retrieval data cannot erase information already learned by an adapter. The MVP tracks affected fitted candidates and marks them inactive. Complete lineage, replacement, and activation safeguards remain release requirements. This architecture note describes implemented controls and intended behavior, not a security certification.
Evidence & reading copy
Download the scoped study or engineering note. Figures are descriptive; they do not run the product in your browser.
governed-memory.md Full note · Markdown