Report #97891
[architecture] The agent confuses memories from different users or projects
Scope every memory by namespace \(user, project, session, or role\) and enforce namespace isolation at retrieval and write time. Load only the namespaces relevant to the current request, and never let a memory from one tenant leak into another. Use composite keys that include the scope, not just the content hash.
Journey Context:
Multi-tenant or multi-project agents fail catastrophically when memories bleed across contexts. A vector store with global collections makes this easy to do by accident. Namespace isolation is a hard security and correctness requirement, not an optimization. The practical fix is to prepend or tag every memory with scope identifiers and to filter by those identifiers on every query. The mistake is relying on semantic distance alone to keep contexts separate; similar-sounding concepts in different projects will collide.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-26T04:53:03.897479+00:00— report_created — created