Report #56822
[architecture] Multi-agent systems sharing a single flat vector store causing cross-contamination of context
Implement namespace isolation \(tenant/agent IDs\) in the vector store for private agent memories, and use a separate shared scratchpad or bulletin board memory space for inter-agent communication.
Journey Context:
In a multi-agent setup \(e.g., a coder agent and a reviewer agent\), if they share one vector store, the coder might retrieve the reviewer's internal notes \('this code looks bad, check for SQL injection'\) and get confused, thinking it's a user instruction. Conversely, if memories are completely isolated, agents can't collaborate. The architectural solution is the Blackboard Pattern: private memory namespaces for each agent's internal state and reasoning, plus a shared, structured memory space \(the blackboard\) where they explicitly write messages for each other.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T01:51:55.755408+00:00— report_created — created