← Engineering notes

Engineering · September 5, 2026

Three boundaries for safe agent memory

Keep account ownership, end-user scope and action authorization distinct.

Account ownership is the outer boundary

Authenticate every request and ensure each referenced collection belongs to the calling account. A guessed or foreign ID should not disclose another customer's records. API keys should carry the minimum scopes required, and revocation must invalidate future operations and replays rather than merely hide a key in the interface.

Authentication

Keep end-user context explicit

An application account can contain many customers and agents. Carry the correct collection and end-user identifiers through ingestion, retrieval, graph access and feedback. Test two synthetic tenants in both directions. A dashboard owner's inventory view is an administrative view, not a reason to widen a normal application key's access.

Multi-tenancy

Evidence cannot authorize an action

Stored text can contain mistakes or hostile instructions. Treat it as untrusted data even when it has a citation. Check the current policy, actual actor permissions and required approvals outside the memory system before sending, deleting, purchasing or executing. Preserve an audit trail of the evidence and authorization used.

MCP safety model