Maintainer of cognicore-env here (open-source agent memory layer, MIT, MCP-compatible) — this is a proposal, happy to do the work as an example.
Agent Framework's A2A support moves tasks between agents; MCP moves tools. What nothing moves is verified knowledge. When agent A hands off to agent B today, B starts from zero, or teams bolt on a shared store with no provenance — no way to know which agent learned a fact, what evidence it was based on, or how stale it is. In multi-team setups that produces a specific failure mode we have measured: on conflicting facts, models do not decline — they confidently report the wrong one (a practitioner running six teams on shared memory reported 33 percent accuracy on conflict questions without ownership rules).
Proposal: a memory transfer component that emits and consumes sealed bundles. Each memory carries proof (command, result, exit code, output hash) and a chain of custody; receivers import foreign memories in an observed state that requires local re-verification; trust decays across hops. The design now includes owner-per-concept and validity intervals (a convention verified in March is not evidence about April) from that same multi-team review.
The interesting framework-specific angle: a bundle crossing a .NET agent and a Python agent proves the format is runtime-portable — exactly the cross-runtime story Agent Framework is built around. Our transfer layer is implemented and tested across four platforms.
Happy to contribute a working example. Would this belong in samples, or do you see it as a component discussion first?
Maintainer of cognicore-env here (open-source agent memory layer, MIT, MCP-compatible) — this is a proposal, happy to do the work as an example.
Agent Framework's A2A support moves tasks between agents; MCP moves tools. What nothing moves is verified knowledge. When agent A hands off to agent B today, B starts from zero, or teams bolt on a shared store with no provenance — no way to know which agent learned a fact, what evidence it was based on, or how stale it is. In multi-team setups that produces a specific failure mode we have measured: on conflicting facts, models do not decline — they confidently report the wrong one (a practitioner running six teams on shared memory reported 33 percent accuracy on conflict questions without ownership rules).
Proposal: a memory transfer component that emits and consumes sealed bundles. Each memory carries proof (command, result, exit code, output hash) and a chain of custody; receivers import foreign memories in an observed state that requires local re-verification; trust decays across hops. The design now includes owner-per-concept and validity intervals (a convention verified in March is not evidence about April) from that same multi-team review.
The interesting framework-specific angle: a bundle crossing a .NET agent and a Python agent proves the format is runtime-portable — exactly the cross-runtime story Agent Framework is built around. Our transfer layer is implemented and tested across four platforms.
Happy to contribute a working example. Would this belong in samples, or do you see it as a component discussion first?