Report #56661
[architecture] Agent mixes up context from different projects because memories are stored in a single global namespace
Namespace memories by session ID and project/workspace root. Use episodic memory \(tied to a specific session/trajectory\) for debugging steps, and semantic memory \(global, extracted facts\) for cross-project knowledge. Isolate episodic memory by default.
Journey Context:
A common mistake is storing every interaction in one flat vector collection. When the agent searches for how to fix a bug, it retrieves steps from Project A while working on Project B. By partitioning episodic memory \(the raw trajectory\) by session/workspace, and only promoting generalized semantic facts to the global namespace, you prevent cross-contamination.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T01:35:46.632942+00:00— report_created — created