Agent Memory · plain Markdown · zero-dependency
Agent Memory Kit — Give Your AI Agent a Memory That Learns Who You Are
Your agent forgets you every session. Most "agent memory" is an opaque vector blob you can't read, locked to one vendor, that only stores — never improves. This is the opposite: a folder of Markdown your agent reads, you can edit, git tracks, and a tiny script keeps clean. It learns from your corrections.
Instant download · no DB, no server, no subscription · Python stdlib + a folder
The problem
Storing isn't remembering
A vector store can find a chunk. It can't tell you what changed, why, or learn that you hate being told the same wrong thing twice.
▌ Typical "agent memory"
- An opaque blob — you can't open it and read what it knows
- Locked to one vendor's API and hosting
- It stores, then stays the same forever
- No provenance — no idea where a "fact" came from
- A monthly bill for a database you don't control
▌ Agent Memory Kit
- Plain Markdown — open it, read it, edit it, diff it
- A folder you own; works with any agent
- Learns from your corrections, gets sharper over time
- Every note is provenance-tagged — you know the source
- Free. Python standard library. Runs offline.
What it does
Four moves, distilled from a real working system
The transferable core of the author's own self-improving memory system — the same loop, packaged to drop into yours.
Save what matters
Durable facts and corrections become provenance-tagged notes — not a blob, readable lines you can trust.
Check memory first
Your agent reads memory before a task. Recall ranks notes by keyword and surfaces the [[wikilinks]] they contain.
Stay sharp, never lose
Weekly cleanup flags exact-duplicate notes and archives thin agent-written ones. It never deletes — it archives, recoverable.
Know who you are
A 15-question quiz becomes a finished CLAUDE.md — how you work, and what you never want to see again.
The maturity ladder
Start where you are. Climb when you're ready.
Most people stop at a static profile file. Levels 2–5 are where memory actually compounds.
The moat: a static CLAUDE.md is table stakes. Levels 2–5 are the part nobody else walks you up — and the kit takes you rung by rung.
What's in the box
Everything you need to run all five levels
- ✓memory.py — the portable tool: capture / recall / curate / status. Stdlib-only, offline.
- ✓15-question Personality Quiz → a finished CLAUDE.md.
- ✓Drop-In Setup MD — one paste into your agent runs the whole setup.
- ✓Build Playbook with the full L1→L5 ladder.
- ✓Illustrated setup guide + the complete honest comparison.
- ✓Free shareable quiz, installer + clean uninstaller (your memory survives).
- ✓Claude Code first — AGENTS.md / soul.md = same pattern, rename the file.
- ✓Obsidian-ready — point it at a folder in your vault; it graphs the [[wikilinks]].
The honest comparison
Where this wins — and where it doesn't
We researched mem0, Letta (MemGPT), Zep, vector RAG, and Claude's native memory. We'd rather tell you the truth than oversell.
Honest: if you need billion-row search or SOC2-on-someone-else's-server, buy that. For one person who wants memory they own, can read, and that gets smarter — this wins.
⚠ Requirements — please read before buying
- An AI coding agent (Claude Code, Codex, Cursor, etc.)
- macOS or Linux — Windows works via WSL
- Python 3.8+ (already installed on Mac/Linux)
- No API keys for the core; Obsidian / Graphify are optional power-ups
Stop re-explaining yourself
Give your agent a memory it actually keeps
A folder of Markdown — capture, recall, curate — that knows who you are. No database. No server. No subscription.
Instant download · 30-day Gumroad refund policy applies
Setting up Claude Code from scratch? Start with the Claude Code Setup Playbook — $29.
See all my Claude Code tools → expressive446.gumroad.com