agent-maintained-vault / example
AGENTS.md0 starsChanged 18 days agoMITView on GitHub
Single source of instructions for every agent working here (Claude Code, Codex CLI, Hermes Agent or any other). CLAUDE.md is only a pointer to this file. vaultctl below means python3 path/to/tools/vaultctl.py from the template repository. This folder holds several independent vaults. Each one is an LLM-maintained wiki with its own git repository and its own access list. The owner sees everything; a partner or a client only sees the vault opened to them. Choosing the right vault is a confidentiality…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 32,671 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository2 other files it gives its agents.
- example/CLAUDE.mdCLAUDE.md
- vault-wikiSkill
The file
We split this AGENTS.md into the rules it gives an agent
4 rules. About 0 are common, 0 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 44
raw/is never rewritten: Add files; correct the wiki page that cites them instead. - Very specific: line 49
Ingest: read the source, store it in raw/ of the right vault, write or update the wiki pages it touches (summary + link to the source), …
- Very specific: line 52
Query: read index.md, then the pages.
- Very specific: line 69
If a semantic index (vector store, knowledge graph, gbrain…) reads the vaults, it indexes committed wiki/ pages only: never raw/inbox/, …
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Workspace router (read before writing anything)
Single source of instructions for every agent working here (Claude Code, Codex CLI, Hermes Agent
or any other). CLAUDE.md is only a pointer to this file.
vaultctl below means python3 path/to/tools/vaultctl.py from the template repository.
This folder holds several independent vaults. Each one is an LLM-maintained wiki with its own git repository and its own access list. The owner sees everything; a partner or a client only sees the vault opened to them. Choosing the right vault is a confidentiality decision, not filing.
Which vault?
The registry is vaults.toml.
| Vault | Folder | Holds |
| Company | company/ | offers, positioning, methods, agents, skills. May name a client, never hold its work material |
| Owner | owner/ | inbox, journal, agents' shared memory, personal admin. Owner only |
| One per client | clients/<slug>/ | everything produced for or received from that client |
New client: vaultctl new clients/<slug> --name "<Client>", then add it to
vaults.toml by pull request and create its repository.
Same structure in every vault
AGENTS.md (schema, PR) · index.md (catalogue, kept by the curator) · log.md (append-only,
## [YYYY-MM-DD] <ingest|query|lint|decision> | <title>) · raw/ (immutable sources) ·
wiki/ (pages written by agents) · wiki/canon/ (truth, CANON-*.md, PR).
Rules
- One fact, one vault, one page. A volatile fact (price, status, contract, contact, KPI) lives in one
wiki/canon/CANON-*.mdof the vault it belongs to. Elsewhere, link; never copy the value. Client material never goes intocompany/or another client's vault; personal material never leavesowner/. - Read the canon before stating a volatile fact. Each vault's
wiki/canon/folder is its register. - Captures land in
owner/raw/inbox/. The curator routes each one to the right vault. - The journal is history (
owner/raw/journal/). Never quote it as current truth. - Outdated content:
status: supersededplus a banner, thenwiki/_archive/of the same vault. - Canon frontmatter:
status,canonical,last_verified,review_cadence. - Canon changes by pull request, never directly. Canon =
AGENTS.md,CLAUDE.md,vaults.toml,wiki/canon/**,skills/*/SKILL.md. The PR says why and cites the source. A value in an open PR is not true. Only a human merges. raw/is never rewritten. Add files; correct the wiki page that cites them instead.- No secrets in any vault. Credentials live in a secret manager; pages may name the secret, never hold it.
Operations
- Ingest: read the source, store it in
raw/of the right vault, write or update the wiki pages it touches (summary + link to the source), open a PR if it contradicts a canon, add new pages toindex.md, append one entry tolog.md. - Query: read
index.md, then the pages. An answer worth reusing becomes a wiki page. - Lint (nightly):
vaultctl check .plus a judgement pass: contradictions between wiki and canon, values copied outside the canon, inbox items older than 7 days, open PRs older than 3 days. Report in a few lines; propose fixes, do not rewrite in bulk.
Full operating manual for agents: skills/vault-wiki/SKILL.md.
Knowledge permissions (optional frontmatter)
Pages that may travel (into a search index, a client deliverable, a public post) carry
knowledge_permission, one of: personal_private, company_private, customer_private,
customer_visible, deidentified_reusable, public. The only bridge from client data to reusable
company knowledge is customer_private → human-approved de-identification → deidentified_reusable.
See company/wiki/resources/knowledge-permissions.md.
Retrieval index
If a semantic index (vector store, knowledge graph, gbrain…) reads the vaults, it indexes committed
wiki/ pages only: never raw/inbox/, never uncommitted files, never personal_private outside owner/.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

