Multi-agent Memory + Identity — 同步／污染／信箱身分

When to use

• CLAUDE.md / rules diverge across Cursor, Claude, Codex, etc.
• Shared memory risks cross-contamination between experiments and mainline.
• Several agents share one inbox/login with no audit of who did what.

Diff vs memory-portability

Problem framing

Public signal: rules files don’t stay in sync; shared agent mail feels identity-less and drops attachments; teams fear one agent’s “notes” poisoning another’s context.

Sync model

1. One memory repo on disk (or git) is canonical.
2. Each tool reads at start; writes only through controlled paths.
3. Rhythm: pull on boot → work → push on end (or timed sync).
4. Conflicts: three-way merge; human resolves SoT conflicts.

Pollution controls

Promotion path: experiment → personal review → project (checklist, not auto).

Identity for shared channels

• Every agent gets a stable identity: display name, From line, signature tag.
• File ownership: creator agent id in metadata or path prefix.
• Ban anonymous overwrite of shared artifacts.
• Audit log: who wrote what, when.

Inbox / artifact protocol (minimal)

id: "..."
from_agent: "money-maker"
to: "human|agent:qa"
subject: "..."
body_path: "inbox/....md"
attachments: ["..."]
status: queued|delivered|dead

• Retries with backoff; dead-letter queue after N fails.
• Never silently drop attachments.

Anti-patterns

• Each tool maintaining its own drifting CLAUDE.md
• Shared folder with no ACL
• Multiple agents on one login with no audit
• Auto-merging experiment memory into project

Acceptance tests

• After tool switch, rules match the memory repo.
• Experiment notes cannot land in project/ without promote step.
• Inbox actions are attributable to an agent id.

Output template

Memory repo path:
Namespaces + ACL:
Sync rhythm:
Agent identities:
Inbox schema path:
Promote checklist:

Out of scope

Building a full multi-agent OS; KYC/email provider signup; publishing without operator approval.