How are you transferring durable agent context without copying the whole local stack?
One practical problem I keep hitting in agent systems is that the useful long-lived context often gets anchored to one machine's local setup.
You can share the prompt. You can share the repo. You can share the tool definitions.
But once "memory" is really a mix of vector state, session carryover, runtime projections, and local machine residue, moving an Agent's learned context becomes much less clean than people imply.
The architecture I've been iterating toward is basically an attempt to stop overloading one storage abstraction with too many jobs. The rough split looks like this:
human-authored policy in files like AGENTS.md and workspace.yaml runtime-owned execution truth in state/runtime.db durable memory bodies under memory/, indexed via MEMORY.md
The important part is not "markdown good, database bad." It's that continuity and durable recall are different jobs. Resume state is about safe handoff between runs.
Durable memory is about procedures, facts, references, and preferences you may actually want to preserve. If those collapse into one opaque local store, "context transfer" often just means "copy the hidden state and hope."
I don't think file-backed memory is a universal answer.
But I do think readable durable memory surfaces make portability less magical and more inspectable. Curious how other people here are handling that boundary. If you actually wanted to move an Agent's learned procedures and references to another machine, where would you want that layer to live?
I'm keeping the repo link out of the body because I'd rather not have this get mysteriously removed as disguised promotion. If anyone wants the full technical framing, I'll put the repo in the comments along with the deeper architecture questions behind it: where policy should live, what should remain runtime-owned, why continuity and durable memory should be separate layers, and what should or should not move across machines.