Context first: I maintain cognicore-env, an open-source (MIT) memory layer for AI agents that runs as an MCP server with no vector database or embedding server. Not proposing we vendor anything — this is an interop idea.
The gap I keep hitting with ADK: sessions and memory services are excellent inside one deployment, but agent knowledge does not travel. Two ADK agents on different runtimes, or an ADK agent and an external agent, share nothing — each re-learns everything. For teams running multiple agents, that is duplicated learning, no provenance, and no way to audit what one agent taught another.
Proposal: a memory export/import hook that emits and consumes sealed memory bundles. Our format carries proof with each memory (the command that was run, the result, the exit code, the output hash) plus a chain of custody across every agent that touched it, and trust decays across unverified hops. So an ADK agent that verified something by actually running a tool can hand that knowledge to another agent with evidence attached, and the receiver treats it as observed-until-reverified rather than trusted by default.
Concrete use case: a code-review ADK agent learns a project's conventions (verified by which PRs merged); an ops ADK agent imports those conventions and stops flagging them as errors.
We have the bundle format and transfer machinery implemented and tested across four platforms, and a design that just went through a public review covering ownership conflicts, validity intervals, and release rules. Happy to contribute an examples/ integration showing ADK memory export/import end to end. What would the maintainers want to see for that?
Context first: I maintain cognicore-env, an open-source (MIT) memory layer for AI agents that runs as an MCP server with no vector database or embedding server. Not proposing we vendor anything — this is an interop idea.
The gap I keep hitting with ADK: sessions and memory services are excellent inside one deployment, but agent knowledge does not travel. Two ADK agents on different runtimes, or an ADK agent and an external agent, share nothing — each re-learns everything. For teams running multiple agents, that is duplicated learning, no provenance, and no way to audit what one agent taught another.
Proposal: a memory export/import hook that emits and consumes sealed memory bundles. Our format carries proof with each memory (the command that was run, the result, the exit code, the output hash) plus a chain of custody across every agent that touched it, and trust decays across unverified hops. So an ADK agent that verified something by actually running a tool can hand that knowledge to another agent with evidence attached, and the receiver treats it as observed-until-reverified rather than trusted by default.
Concrete use case: a code-review ADK agent learns a project's conventions (verified by which PRs merged); an ops ADK agent imports those conventions and stops flagging them as errors.
We have the bundle format and transfer machinery implemented and tested across four platforms, and a design that just went through a public review covering ownership conflicts, validity intervals, and release rules. Happy to contribute an examples/ integration showing ADK memory export/import end to end. What would the maintainers want to see for that?