Design review: how should memory be isolated between orchestrators?

[Weegy]

This is the design tension I'd most like external eyes on.

Today: strict per-orchestrator memory namespaces + knowledge-graph slices, so context can't bleed between agents. Clean — but it makes deliberate cross-agent sharing the awkward case.

Open questions:

1. Where should the boundary live — per-agent, per-team, capability-scoped?
2. For tool output: sanitise before or after record expansion?
3. How do you model intentional sharing without reopening the leak?

How have you solved this in your own multi-agent setups?