Verified Memory Vault

Free · Obsidian vault · Persistent memory for agents that proves it works

A small vault that gives AI coding agents (Claude Code, Codex, Gemini CLI, …) a persistent memory — and, unlike a bare folder structure, it checks itself and protects itself:

• memory_check.py scores the memory's health at any time: protocol violations, dead links, bloat, inbox pressure. One command, honest report.
• memory_guard.py (optional git hook) refuses the commit that deletes your agent's brain — the documented way agent memories die.

Companion to the human-side Second Brain Starter — the free starter vault for people who want a working note system in 10 minutes. https://secondbrainstarter.github.io/

---

Why this exists

"Give your agent an Obsidian vault as memory" is now a known idea. What breaks in practice is not the idea but the maintenance:

1. Memory rots — undated entries, duplicates, and speculation fossilize, and a bloated MEMORY.md quietly eats the context window.
2. Memory dies — agents occasionally delete or rewrite their own memory files while "cleaning up".
3. Nobody notices until it's too late — a folder structure cannot tell you whether it still works.

This vault addresses all three with two dependency-free Python scripts and a strict protocol. No Node, no TypeScript, no plugins, no database — if you can run python3, you can verify your agent's memory.

What's inside

• CLAUDE.md — boot file every agent reads first (rules + protocol)
• MEMORY.md — durable memory: dated, append-only, one line per fact
• 01_Daily/ — one note per session: work done, decisions, blockers
• 00_Inbox/ — capture everything; sort later
• 02_Templates/ — daily note templates (EN + DE)
• tools/memory_check.py — health check (score, problems, exit code). The score is telemetry for humans; the exit code is the contract, graded by risk class: 0 = healthy, 1 = degraded (hygiene defects like dead links or bad note names), 2 = provenance breach (undated, duplicated, or near-duplicated memory entries — the memory can no longer be trusted about when something was learned). Near-duplicate detection compares date-stripped, stopword-filtered token sets: a fact re-captured in different wording or with an added detail is flagged even though the lines are not byte-identical. Gate session-shutdown hooks or CI steps on it; treat exit 2 as strictly worse than exit 1.
• tools/memory_guard.py — git pre-commit guard against mass deletion (file count AND absolute removed-lines cap) and history rewrites of MEMORY.md

Quick start (10 minutes)

1. Install Obsidian (free): https://obsidian.md
2. Download the vault: grab the ZIP from the latest release — or clone this repo.
3. Unzip → Obsidian → "Open folder as vault". Daily notes and templates are preconfigured.
4. (Recommended) Make it a git repo and install the guard: git init && ln -s ../../tools/memory_guard.py .git/hooks/pre-commit
5. Tell your agent to read CLAUDE.md at session start and to run python3 tools/memory_check.py once per session.
6. End of session: fill the daily note, promote durable facts to MEMORY.md. That's the whole habit — and now it's verifiable.

Gate automation on exit codes, not scores: memory_check.py returns 0 (healthy), 1 (degraded — hygiene defects) or 2 (provenance breach — undated/duplicated/near-duplicated memory entries, the stricter class). Hook that into session shutdown or CI; the printed score is a dashboard for you, not a contract for machines.

Verify your copy in one command (no dependencies, stdlib only):

python3 tests/self_test.py

It builds throwaway vaults and real throwaway git repos, then asserts both contracts end-to-end: healthy → 0, hygiene defect → 1, undated/duplicated/near-duplicated → 2 — plus every memory_guard refusal (mass deletion, history rewrite without archiving, removed-lines cap) and the legitimate paths (append, archive-and-trim) staying allowed. The same suite runs on every push via GitHub Actions (latest run).

The three rules

1. Capture, don't sort. Everything lands in 00_Inbox/ first.
2. Every session leaves a daily note. So the next session resumes cold.
3. Promote durable facts to MEMORY.md. Append-only, dated, one line, verifiable only.

How this differs

• vs. plain vault templates: this one checks its own state instead of assuming it.
• vs. heavy agent-memory frameworks (hooks, MCP servers, semantic search): everything here is readable Markdown plus two ~200-line stdlib scripts — you can audit all of it in ten minutes.
• Bilingual by design (EN + DE); the agent mirrors your language.

License

Vault content: CC BY 4.0 — free to use, including commercially; attribution required (LICENSE.md). Tools under tools/: MIT.