Claude Code Guide: Persistent Memory Playbook for AI Agents, Hooks, Watchdog, Code Kit (PDF)

Give your coding agent a memory that survives the end of a session.

A field guide from a memory system that has been running every day for
three months: one fact per file, an index that stays small, a recall hook
that opens the right file at the right time, sessions that talk to each
other, and tests that keep all of it honest. Written for people who use
Claude Code (or a similar agent) and are tired of explaining the same
things every morning.

WHAT IS INSIDE (24 pages, 12 chapters)

1 Why your agent forgets - and why a bigger context window does not fix it.
2 The neuron: one fact per file. Frontmatter that does work, the four
types, the rules that came from pain.
3 The index: a map that must stay small. The byte budget, and the day the
index truncated from the end, silently.
4 Recall: the right neuron at the right time. A dependency-free BM25 recall
script printed in full (70 lines of Node.js), when to add embeddings, the
cold-model tax, and the measurement that fooled us first.
5 Episodic and semantic: two stores, two laws. The checkpoint that survives
compaction.
6 Hooks that make it automatic: SessionStart, UserPromptSubmit, Stop and
PreCompact, PreToolUse - with the settings block.
7 Many sessions, one brain: session mail, a sleeping watchdog that costs
nothing while it waits (exit codes, lock, liveness), a sentinel for silence.
8 Keeping it honest: gates, sabotage, green-for-the-wrong-reason, the four
questions of a lock.
9 Fourteen pitfalls we paid for.
10 Twelve repositories worth knowing - stars and licences verified against
the GitHub API on 12 September 2026, with our verdicts.
11 Quick start in an afternoon: eight steps and the templates.

Every number in the book was measured on a working system. Where we were
wrong, the book says so.

THE CODE KIT (ZIP)

Every script printed in the book, ready to run - no packages to install:
recall-bm25.js (lexical recall that returns pointers, never verdicts),
recall-hook.js (the UserPromptSubmit wrapper), session-start.js and
checkpoint.js (the SessionStart and Stop/PreCompact hooks), watchdog.sh,
a seven-gate self-test that deletes a neuron and checks recall loses it,
the hook settings example, a five-neuron sample memory with its index, and
the neuron and index-line templates. Requires Node.js 18+; the watchdog
needs bash. The kit was extracted from a fresh copy and executed before
release.

WHAT YOU NEED

Any PDF reader. The book is vector, with embedded fonts, in A4 and US
Letter. No account, no subscription, no API key.

WHAT YOU GET

An instant download: the PDF in two page sizes, the code kit, and a short
read-me. Nothing is shipped. The link appears as soon as your payment clears.

The code kit is also public on GitHub (MIT, with a self-test that runs on
every commit): github.com/yasinkcr/claude-code-persistent-memory. You are
paying for the book - the 24 pages that explain why each piece is shaped
the way it is - not for the scripts. Read the code first if you like.

HOW THIS WAS MADE

Written with AI assistance, from a system we built and run ourselves. The
chapters, the scripts and the checks were designed deliberately - not
generated from a prompt. Every listing is checked against the real GitHub
API; every script is run from the packaged ZIP before release.

TERMS

For your own personal use. You may not resell, redistribute or share the
PDF itself. The code in the kit may be used and modified freely in your own
projects (no warranty).

Digital products cannot be returned - but if something is unclear or a
script does not run for you, message me and I will fix it.

Not affiliated with or endorsed by Anthropic. "Claude" and "Claude Code" are
trademarks of Anthropic, PBC, used here only to describe compatibility.