A Free LLM Wiki Template for Claude Code, Built From Six Months of Real Use

This is the product page for the free personal-memory template: an LLM wiki template you can copy into a private repo and grow with Claude Code.

The architecture is the five-layer Claude Code memory system from March. After running that setup daily since spring, I audited it on 22 October 2026 — about 860 markdown files by then — and wrote up seven failures and the rules that fixed them. This post is the template those rules shipped in: folder tree, /setup, the facts-vs-skills split, lint.py, and the no-code variant. Examples use Acme and Example Client, never real clients.

The measured before/after numbers below describe my knowledge base, not the empty public scaffold. The template ships the rules and the limits.

👉 Use this template on GitHub — click Use this template, make the copy private, run /setup. Star the repo if it helps you keep the pattern around. Product page: /products/personal-memory-template/.

What you get (folder tree)

The full variant is a normal git repo. The layout matches what the public README shows:

personal-memory/
├── MEMORY.md          ← routing index (≤150 lines)
├── AGENTS.md          ← the one rules file (CLAUDE.md only imports it)
├── ORG.md             ← human review routine
├── log.md             ← this month's changes; older months in archive/log/
├── CHANGELOG.md       ← skill changes (Keep a Changelog + SemVer)
├── .claude-plugin/    ← plugin.json: version number for the skill set
├── skills/            ← setup, ingest, invoice-client (versioned)
├── clients/           ← entry file per client + a folder when it grows
├── companies/         ← your own companies
├── engineering/       ← patterns + debugging/ (one file per topic)
├── playbooks/  policies/  home/  projects/
├── scripts/lint.py    ← health check, standard library only
├── enkel/             ← simple Norwegian variant (Claude Project / Notion)
└── archive/           ← ignored by agents

MEMORY.md routes. Topic and client files hold the facts. AGENTS.md is the always-on rules file (agents.md format); CLAUDE.md is only @AGENTS.md, the same pattern as in CLAUDE.md too long. archive/ is ignored unless you ask for it.

Start in five minutes with /setup

1. On GitHub, click Use this template (or clone the repo) and make your copy private — it will hold client data.
2. Open Claude Code in that folder.
3. Run /setup. It checks the repo is private, interviews you one question at a time (clients, invoicing, systems, recurring work), then creates the first client and company files, the index, and starter skills.
4. Point your global ~/.claude/CLAUDE.md at the repo’s MEMORY.md so other projects can find the wiki.
5. After finished work, run /ingest so what you learned lands in the right files. That ingest idea is the same loop as in Claude Code automation, scoped here to the knowledge base.

You do not start from a blank folder and invent the schema. /setup builds the first version; you edit from there.

Facts in files, procedures in skills

Two jobs, two homes:

• Facts live in markdown files — clients/acme.md, quirks, deploy notes, decisions for Example Client.
• Procedures live in skills — skills/invoice-client/SKILL.md, skills/ingest/SKILL.md.

A client file points at a skill by name (for example invoice-client). The skill reads the client file. Neither copies the other. That split is why the rules file can stay short: always-on behavior in AGENTS.md, reference material in the wiki, steps in skills. Anthropic’s Agent Skills best practices push the same direction — concise skill bodies, clear triggers, detail loaded on demand.

Durable facts do not belong in Claude Code’s project auto-memory folder. That index has a documented load window (200 lines / 25 KB); when mine grew past it, the newest memories never reached a new session — full write-up: Claude Code auto memory not working?.

The guardrails: size limits and lint.py

Limits exist because the files that grow past them stop being useful. In my own knowledge base (22 October 2026 audit):

The template’s AGENTS.md encodes the caps the lint script enforces: MEMORY.md ≤ 150 lines (rows ≤ 150 characters), client entry files ≤ 120 lines, topic files flagged over 40 KB, log.md kept to the current month.

Run the check yourself:

python3 scripts/lint.py

It reports broken links, oversized index and entry files, skill version gaps, and related warnings. In my knowledge base a hook runs that script before every commit (dead links 31 → 0). The public template does not ship a pre-commit hook or a changelog generator — only scripts/lint.py and a CHANGELOG.md you maintain. If you want the check on every commit, add a git hook yourself. The README’s standing advice is still: run lint monthly and fix what it finds.

Versioned skills: ingest and invoice-client

The template ships three skills under skills/:

• setup — the /setup interview that builds the first wiki.
• ingest — after finished work: update the right topic files, one line in log.md, then lint.
• invoice-client — example domain skill: how to build an invoice draft for a client; it reads clients/<slug>.md instead of hardcoding Acme’s rules.

Each SKILL.md has a version line. Reading maps name the skill, not the version, so bumping a skill does not force edits across every client file. .claude-plugin/plugin.json holds one version number for the set; CHANGELOG.md records what changed.

In my own knowledge base, skills with a version went from 30 of 64 to all 64 in that October audit. Search recall on 20 test questions (right file in the top 3) went from 50% to 85% after the broader cleanup — including clearer files and better matching — not from the template alone.

No code? The simple variant for Claude Projects or Notion

If you do not want a git repo, use the enkel/ folder: a small Norwegian variant for a Claude Project or Notion. Same idea — one file (or page) per client for facts, a skill for the procedure — without lint.py or a routing index in the Claude Code sense.

• Claude Project: upload the client files and project instructions; re-upload when something changes.
• Notion: one page per client, connect Notion to Claude, keep the same facts-vs-steps split.

For a full personal task system in Notion (inbox, next actions, weekly review), that is a different job: GTD Notion setup for 2026. The simple wiki variant is only the client knowledge layer.

How it differs from other LLM wiki templates

Karpathy’s LLM Wiki gist is the pattern this repo follows: immutable raw sources, a compiled wiki, a schema for how the wiki is written, plus ingest / query / lint as operations. Most generic “second brain” dumps stop at a folder of notes.

What this template adds on top of that pattern:

1. One rules file (AGENTS.md) with a Claude-friendly import, instead of drifting copies per tool.
2. Hard size limits checked by scripts/lint.py, not only documented as advice.
3. /setup so the first version of the wiki is interviewed into existence.
4. Versioned skills with facts kept out of the procedure files.
5. A simple non-git variant for Claude Projects or Notion.

It is a Claude Code knowledge base (and a Claude Code second brain, if you use that phrase) shaped by daily client work — not a RAG demo and not my private 860-file tree. Clone the empty scaffold; bring your own Acme and Example Client facts.

👉 Use this template on GitHub — Use this template, make it private, run /setup. Star the repo if you want it easy to find later.

Series context: architecture in the March memory system post, measured failures in 7 lessons, then the CLAUDE.md trim and auto-memory load window. This post is the template itself.

Subscribe below if you want template updates and later series posts by email.

FAQ: LLM wiki template