POC2 · Agent memory with history, across two runners P2
Write phase (per role). ruflo memory init, memory store a lesson in the role's namespace, memory list to a text file. Clone the memory vault (key on stdin), create branch mem-<role>-<run>, copy the durable stores (.swarm/memory.db, .swarm/agentdb-memory.db, .claude/memory.db, ruvector.db) and the list under memory/<role>/, commit, push, switch to the default branch, merge, push.
Read phase (fresh runner). Clone the vault into an empty directory, copy each role's files into a fresh folder, run ruflo memory list -n <role>. Same entries come back.
ROLE=coder bash poc/poc2-agent-memory-in-vault/run.sh write
ROLE=tester bash poc/poc2-agent-memory-in-vault/run.sh write
rm -rf .vault-memory && bash poc/poc2-agent-memory-in-vault/run.sh read
sgit history log --oneline -n 10 # inside .vault-memoryAcceptance. Two writers, one fresh reader, same answer; history has a commit per role; the server never sees plaintext (sgit encrypts client-side, AES-256-GCM). Met locally; in CI the poc2-write matrix and poc2-read jobs run on separate runners once the secrets exist.
Why it matters. Per RuVector's own README (“Security and governance”): its witnesses and hash-linked logs “do not encrypt memory content”; “collections and metadata filters organize memory; they are not a complete authorization boundary”; and “a delete from the live store does not automatically remove copies in snapshots, branches, replicas, exports”. RuVector already has snapshots and copy-on-write branching (rvf derive). The vault complements them: encryption on the client, a history shared across machines, and a vault that is deleted with the run.
Output: the fresh runner's answer
loading…Output: vault history (what was learned, when, by whom)
Evidence: evidence/poc2-run.txt, evidence/poc2-history.json. Two things we learned: a vault with zero commits cannot be cloned (seed it with a README), and sgit's default named branch is called current, not main.
