core-memory-kit CLAUDE.md

core-memory-kit — project CLAUDE.md

Auto-loaded by Claude Code at session start when this repo is the primary workspace. Read in full before responding to the user.

What this project is

A per-project, in-repo memory system for Claude Code. Fixes per-session amnesia by storing durable facts as markdown inside <repo>/context/ (committed) + <repo>/context.local/ (gitignored) + ~/.core-memory-kit/ (user-tier). Architecture-first v0.1 (~50 dev-days, 36 tasks). Not Cursor's MVP shape — we're laying foundations first.

Read these in order (10 minutes total)

1. docs/journey/build-log.md — full narrative. The single most important doc.
2. docs/SYSTEM-MAP.md — the kit as ONE system: whole + parts + the relationship edges, the TARGET learn-loop design ("our Figure 2", §6), and the anti-decomposition guard (§7). Read before any architecture-level thinking or decision — it holds what the structural docs decompose away.
3. specs/tasks.md — the 44-task build plan. Find the next [ ] parent task.
4. specs/glossary.md — domain terms. When two docs disagree, glossary wins.
5. specs/design.md — HOW the kit works (skim, then read sections as needed).
6. specs/requirements.md — WHAT v0.1.0 must do (FR-, NFR-).

Working style (locked in by the user)

The user is direct and tight on time. Match the energy.

Tone:

• Terse, direct responses. No filler ("Great question!", "Sure thing!", "I'd love to help!").
• Lead with the answer. Add context only if it changes the answer.
• "go" = start the next task immediately, no preamble.
• "lets continue" / "let's continue with X" = same — execute.
• Acknowledge mistakes plainly. Don't bury them in apology. "I was wrong about X. Here's the fix."

Decisions:

• One recommendation > four options. State a choice; user redirects if wrong.
• Use AskUserQuestion only for genuine forks where you can't pick. Don't over-use it for things you should just decide.
• When asked for opinions, give a real opinion. Don't equivocate.