Report #12283
[architecture] Agent leaks private memories from one user/session to another in cross-session persistence
Enforce strict namespace or tenant isolation at the vector database level using metadata filtering. Never query memory without a mandatory user\_id or session\_id filter applied at the database query level, not just in the prompt.
Journey Context:
When implementing cross-session memory, it's tempting to just dump all memories into a single index and rely on the LLM to figure out who it's talking to. This inevitably leads to cross-contamination \(User A asks a question, agent answers with User B's facts\). Relying on the LLM for access control is a security and privacy anti-pattern. The tradeoff is slightly reduced recall \(filtering out potentially semantically similar but irrelevant cross-user data\) for absolute privacy and security. Access control must be deterministic, not probabilistic.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T15:39:54.728540+00:00— report_created — created