[Discussion] Preventing memory leaks between agents via tenant isolation
#28
by Spy9191 - opened
When scaling from single-agent to multi-agent architectures, flat knowledge graphs frequently leak context. If Agent A stores an observation, Agent B can accidentally retrieve it via semantic similarity if there is no hard boundary at the storage layer.
The architectural fix requires enforcing tenant_id and agent_id boundaries directly in the memory store, treating cross-agent context sharing as an explicit policy rather than a default retrieval behavior.
We recently implemented this isolation natively in Smriti using partitioned schemas. How are developers in the HuggingFace ecosystem handling memory isolation in lightweight multi-agent deployments?