Let's get down to business. How do you keep shared agent memory honest?
My agents share one repo of markdown notes as their memory. Writing to it is easy. Keeping it true not so easy: one stale "ready to send" turned into four wrong documents in ninety minutes.
Commit
e040c63
i use brigade, agents can share memory across codex, claude, openclaw, dot, grokbot , etc.. and agents leave a verification record of what they did with exit codes, linease, run journal and all that fun jazz.
when you work with multiple agents they can pass each other evidence…
Evidence over trust is where I've landed too. One of my rules: a handoff from another agent is a guess until something proves it. When an agent receives an evidence package, does it re-check anything, or take the exit codes as proof?
commit e040c63
it re-checks. receipts are pinned to the tree fingerprint they ran on, so if the code moved since the handoff, the old exit codes don't count and it reruns. the part it can't verify is a judgment like "ready to send." those stay a guess until something re-proves them on the…