Skip to content

Shared agent memory

A newly mounted shared folder is hundreds of opaque files to an agent. It won’t know what lives where, where to write, or that the folder exists at all unless something tells it.

Two files fix that. They have different jobs, and you need both.

<shared>/AGENTS.md — synced, team-wide.

The single source of truth for conventions: what each area is for, naming patterns, where agents should write (reports → reports/), what not to touch.

Because the folder syncs, the map travels with it. Every member on every platform gets it, and hub history tracks who changed the rules.

# Team knowledge folder
- `reports/` — generated reports. Agents write here. One file per topic,
kebab-case, dated in frontmatter.
- `decisions/` — architecture decision records. Append only; never rewrite
an existing ADR.
- `runbooks/` — operational procedures. Keep them executable.
- `inbox/` — unsorted. Fine to read, don't build on it.
No secrets in this folder — anything here can be shared as a public link.

Keep it under a screen. It is scaffolded once, by the project creator — joiners read it and follow it. A new member’s agent must not rewrite team conventions on day one.

Per machine, never synced.

For --shared mounts inside a repository, append two or three lines to the repo root’s AGENTS.md and/or CLAUDE.md (both, if both exist):

## Shared knowledge folder
`wiki/` is synced across the team via BearDrive. Read `wiki/AGENTS.md` before
working there. Put shareable artifacts there so teammates' agents see them.
No secrets.

Point at the synced map. Don’t duplicate its conventions, or the copy goes stale.

Why the pointer isn’t optional politeness

Section titled “Why the pointer isn’t optional politeness”

Platform discovery differs:

Platform Finds <shared>/AGENTS.md on its own?
Claude Code / Cowork Lazily — loaded when a file in that subtree is first read
Hermes Lazily — progressive discovery; walks up from files it touches
Codex Never — only loads AGENTS.md along the root→cwd path

And even where lazy loading works, it fires only after the agent decides to enter the folder. Only the root pointer gives it the awareness to go there in the first place — “save the report where the team sees it.”

/beardrive:install offers to write both files, each as its own consent. Never write either silently into someone’s repository.

Worth teaching your agents as a habit: on first contact with a synced folder, read its AGENTS.md before substantive work.

If there is none, orient from the file tree plus recent activity —

Terminal window
bdrive log wiki # which areas are alive

— and, if this device created the project, offer to draft AGENTS.md for the team.

Good candidates are the things that are expensive to rediscover and cheap to write down:

  • Findings — what an investigation concluded, and what was ruled out.
  • Decisions — what was chosen and why, so the next agent doesn’t relitigate.
  • Runbooks — procedures that worked.
  • Artifacts — reports, analyses, and generated documents teammates need to read.

What doesn’t belong: secrets and credentials (any member can mint a public link for any file), anything derivable by reading the code, and transcript-level detail nobody will ever read again.

See Scoping the folder for keeping build output and dependencies out mechanically.

Stamp changes with the context that produced them:

Terminal window
bdrive sync --note "session 8f21: auth refactor investigation"

The note shows up in bdrive log and hub history, and keeps applying to daemon-committed changes until --note-ttl (default 30m) expires. This is how a change traces back to the conversation that caused it.