Cdb Scan

Map this project into memory

Everything else in claude-db's memory is testimony: it records what happened, turn by turn, as it happened. A fresh install knows none of that, and stays useless for weeks while it fills up.

This fills the gap from the one source available on day one — the code itself, in two passes.

Pass 1 — the code graph

Run this first, from the project root:

claude-db scan

It parses every supported source file and stores each symbol and each relationship between them: what calls what, what imports what, what extends what. It is deterministic, costs no tokens, and takes seconds. Report the symbol and edge counts it prints.

Once it has run, find_usages answers structural questions from the graph:

• mode: "usages" — what references this symbol, with the relation on each line
• mode: "explain" — that, plus what the symbol itself reaches
• mode: "path" with target — how two symbols connect

Prefer those over reading files when the question is about structure.

Pass 2 — the written profile

The graph records what the code is. It cannot say why the project is built this way, so the sections below capture that in prose. What you write here is your reading of the codebase, not a record of events, so it is tagged inferred and must never be phrased as history.

Work through the sections below in order. For each one, look at the actual files, then call the remember MCP tool once with the given key.

key is what makes this re-runnable: writing the same key again replaces that note rather than adding a second copy. Use these exact keys, and always pass tags: ["inferred"] and kind: "context".

Before you start, call search for profile in this project. If notes already exist, read them first and update what changed rather than restating it — a second run should be an edit, not a rewrite.

1. profile:stack

Languages, runtime and version floors, frameworks, package manager, database, test runner, build tool. Read the manifest (package.json, pyproject.toml, go.mod, Cargo.toml) and the lockfile rather than guessing. Name versions where a version constrains what can be written.

2. profile:layout

The directory map, one line each, and what lives where. Include the entry points: the binary, the server, the main export. Skip anything generated.

3. profile:conventions

How this codebase is written, taken from the code and not from a style guide: module system, import style, error handling, naming, comment density, how tests are structured. Note where the project deviates from the language default, since that is what someone would otherwise get wrong.

4. profile:workflows

The commands that matter — install, build, test, lint, run, release — and anything non-obvious about running them: a required service, an environment variable, a step that must come first. Take them from the manifest scripts and CI config, not from the README's aspirations.

5. profile:architecture

Only the parts a newcomer could not infer from the layout: the main abstractions and how a request or a command actually flows through them. Two paragraphs at most. If the codebase is small enough that the layout says it all, skip this section rather than padding it.

Rules

• One remember call per section. Five notes, five stable keys.
• Write claims, not headings. "Storage is one MemoryStore interface with three adapters, selected by URI scheme" — not "Storage architecture".
• Say what you verified. If something was not read, leave it out. An inferred profile that is confidently wrong is worse than a short one.
• Never restate CLAUDE.md. That file is already in the system prompt on every turn. Memory is for what is not.
• Do not record secrets, credentials or personal data, even if the code contains them.