Every AI coding agent starts every session with amnesia. The repository is
there, the instructions are there — but the knowledge from yesterday's session
is gone: the workaround for that build error, your preference about naming,
the decision you made last week. So you explain it again. And again.
The fix is almost embarrassingly simple: give the agent a place to write things down — and a way to check that the memory is actually healthy.
CLAUDE.md
at the start of every session.What is inside:
00_Inbox
— everything lands here first (capture, don't sort)01_Daily
— one note per session: what was done, decided, blocked02_Templates
— daily note templates, English and GermanCLAUDE.md
— the boot file every agent reads first: rules and memory protocolMEMORY.md
— the durable memory: append-only, dated, one line per facttools/memory_check.py
— scores memory health on demand: protocol violations, dead links, bloat, inbox pressure. One command, honest report.tools/memory_guard.py
— an optional git pre-commit hook that refuses the commit that deletes your agent's brain (the documented way agent memories die)The whole method is three rules:
Capture, don't sort. New notes land in the inbox; sorting happens in the daily note, not while capturing.
Every session leaves a daily note. So the next session resumes cold — without asking you where you left off.
Promote durable facts into MEMORY.md. If a fact will matter in a later session, it belongs there: dated, one line, never rewritten.
Why "self-checking" matters: folder structures rot silently. Memory gets bloated, links die, entries lose their dates — and nobody notices until the agent's context is full of junk. Run
python3 tools/memory_check.py
and you get a score, a list of concrete problems, and an exit code your CI or agent loop can act on. If you keep the vault in git (recommended), the guard hook blocks mass deletions of MEMORY.md
before they happen.
Setup takes ten minutes. Install Obsidian (free), download the vault,
open the folder as a vault — daily notes and templates are already
preconfigured, no plugins to install. Claude Code reads CLAUDE.md
automatically; for other agents (Codex, Gemini CLI, …) point them at the file
or copy it to their boot file, e.g. AGENTS.md
. At the end of a session:
fill the daily note, promote the durable facts, run the check once. That is
the whole habit.
And what it is not — honestly: not a plugin collection
(no community plugins, no vector database), not an autonomous system (it is a
structure with two small scripts, not a magic product), and not a
replacement for your code repository — it is memory about the work. That is
exactly why it keeps working: nothing to maintain, everything auditable in
ten minutes.
Why the license matters: many “AI vaults” in circulation are
non-commercial. This one is licensed CC BY 4.0 — free to use,
including in your own business; only attribution is required. If you already run
a note system for yourself, Second Brain Starter is the human-side companion: the same three-folder philosophy, built for people who want a working note system in ten minutes.
Download Verified Memory Vault v0.9 (ZIP)
Source and updates: github.com/secondbrainstarter/verified-memory-vault
Question, feedback or an idea? Write to geld.hamster@gmx.net — every mail is read.