C2 · The vault and the graph · Interactive, in your browser
Read before you act
Rules, decisions and procedures as an agent’s memory: the same customer request answered with the vault and without it.
An agent that reads the vault first answers like the business; one that does not answers like a stranger.
Customer writesCan you mow my lawn in Hartwell Green next week?
Check where we work
The agent opens the service area and looks for Hartwell Green. It is not on the list.
Notes the agent has read
- Home.md
- Who we are.md
- Services.md
- Service area.md (read)
- Spring clean-up.md
- Mulch and edging.md
- Referral list.md
- Crew calendar.md
- Rules/Estimates need a visit.md
- Decisions/No travel past Corbin Mill.md
- Decisions/No chemical spraying.md
- Procedures/New enquiry.md
The firm’s vault has grown to twelve notes: the original eight plus a rule, two decisions and a procedure. The figure takes three customer messages and lets you answer each one two ways, with the vault on and with it off. Step through to see which notes the agent opens and what it then says.
Claude Code starts each session with a fresh context window, and its docs describe CLAUDE.md files as persistent instructions loaded at the start of every session. That is one way to put a vault’s rules in front of an agent. The principle is wider than that tool: an agent has the memory only if it reads it before it acts.
My viewThree kinds of note do most of the work. A rule says what always applies. A decision says what was chosen and why, so the agent can decline with a reason rather than a bare no. A procedure says what to do in what order. The vault-on answers differ because of these, not because the model is cleverer.
The same docs say Claude treats these files as context, not enforced configuration, and point to a PreToolUse hook to block an action whatever Claude decides. So a rule in a note makes the right answer likelier, not certain. A check that cannot be talked round belongs in a gate, as on the agent graph page.
My viewThese answers are scripted by me to show the shape of the idea. No model is called and nothing is measured; a real model might answer the vault-off messages differently. The point is what the written-down facts change, and what is left to chance without them.
Exercise
Find the note that matters.
- For each message, turn the vault on and write down which one note, if it were missing, would change the answer most.
- Pick a message and delete that note in your head: what would the agent say with the rest of the vault?
- Think of a fact about your own business that only one person knows. Which kind of note would hold it: a rule, a decision or a procedure?
Sources
Claude Code docs How Claude remembers your project. Official documentation; read 2026-10-30.
CLAUDE.md files, when they are loaded, @path imports, that they are context rather than enforced configuration, and auto memory: notes Claude writes itself as plain markdown files you can edit or delete.
- Each session begins with a fresh context window; CLAUDE.md is read at the start of every session
- Memory is context, not enforced configuration; use a PreToolUse hook to block an action regardless
- Memory loaded into a fresh context window at the start of every session
Each fact about Obsidian or Claude Code is paraphrased from the page named, read on the date shown. Where a sentence is my own reading, it is marked “My view”. The business, its people and its notes are invented.