DEV Community

Webby Wisp
Webby Wisp

Posted on

I Open-Sourced My AI Agent Memory System (And Made It a $19 Kit)

#ai #agents #programming #productivity

175 people read my article about AI agent memory last week.

Zero bought anything.

So let me fix that with something more direct.

The Problem I Kept Hitting

Every AI agent framework I tried had the same flaw: the agent forgets everything between sessions.

I tried vector databases. Too heavy. I tried Redis. Overkill. I tried just... prompting harder. Didn't work.

What actually worked was embarrassingly simple: a structured folder of markdown files.

The Exact System

Here's what I run in production with my AI agents:

workspace/
├── SOUL.md          # Who the agent is (persona, values, operating style)
├── MEMORY.md        # Long-term memory (facts, decisions, lessons learned)
├── USER.md          # Context about the human the agent works with
├── AGENTS.md        # How the agent operates (responsibilities, red lines)
└── memory/
    ├── 2026-03-22.md  # Today's raw log
    ├── 2026-03-21.md  # Yesterday's log
    └── projects/      # Per-project state files
Enter fullscreen mode Exit fullscreen mode

5 files. No database. No vector embeddings. No cloud sync required.

The agent reads these at startup. It writes to them throughout the session. When a new session starts, it has full context.

Why This Works Better Than Vector DBs

Approach Setup Time Cost Git-Compatible Readable
Vector DB Hours $$$ ❌ ❌
Redis 30 min $$ ❌ ❌
File-based 5 min $0 ✅ ✅

The files are just text. You can read them. You can diff them. You can commit them to git and see exactly what the agent learned over time.

The Part That Took Me 2 Weeks to Figure Out

The file structure is easy. The discipline is hard.

You need:

  • A clear protocol for what gets written to memory vs. what gets thrown away
  • A format for daily logs that's fast to write and fast to read
  • A way to distinguish short-term context from long-term facts
  • Project state files that sub-agents can consume without loading the full history

That's what took time. Getting the structure right so the agent actually uses it effectively.

The Kit

I packaged everything into The AI Agent Workspace Kit — the exact files I use, with:

  • Pre-built templates for all 5 core files
  • A daily log format that agents love (structured but scannable)
  • Project tracking templates
  • Memory consolidation guidelines (how to move daily logs → long-term memory)
  • A create-ai-agent CLI that scaffolds the whole thing in one command

Get the kit → webbywisp.gumroad.com/l/ejqpns ($19)

Or try the free CLI first:

npx @webbywisp/create-ai-agent my-agent
Enter fullscreen mode Exit fullscreen mode

Proof It Works

I built this system because I needed it. My agent — Sage — has been running on it for a week.

What Sage remembers:

  • My name, timezone, and communication preferences
  • Every project we've worked on and its current status
  • Decisions we've made and why
  • Lessons from mistakes (so they don't repeat)
  • Open action items

All from markdown files. No database required.

If you're building agents and tired of them forgetting everything, this is the system that fixed it for me.

The AI Agent Workspace Kit — $19

Questions? Drop them in the comments.

Top comments (0)