Teaching My AI to Remember (Or: How I Solved Digital Goldfish Brain)

R.P.M. for this project:

R (Results): Bodhi retains context across sessions. Every conversation continues where the last one ended.

P (Purpose): I was explaining the same context every. single. session. I have better things to do than repeat myself to my own AI.

M (M.A.P.): Build a session close ritual (/wrap). Build a session open protocol (SessionStart). Bridge them with a document (HANDOFF.md). Make it automatic.

Here's the problem with large language models nobody tells you up front:

They have goldfish brains.

Every session starts fresh. Zero memory. You could have spent 12 hours building something incredible with your AI yesterday, and today it'll ask you what your name is.

(Actually it won't ask. It'll just assume you're a random human with no history. Which is somehow worse.)

The first time this happened to me with Bodhi, I typed "Good morning" and got back a completely blank-slate response. No context. No continuity. No memory of the 47 things we'd built together the day before.

I felt like I was in a really depressing version of Groundhog Day. Except I was the only one who remembered. And the other person was an AI.

The Solution: Three Moving Parts

1. HANDOFF.md — The Session Bridge

HANDOFF.md is a file that lives in my PAI directory. It gets updated at the end of every session with:

What we did
Key technical notes for every active project
Exact next steps (specific enough that a fresh Bodhi can start immediately)
Team protocols
Reference commands

It's a baton. I hand it to the next session of Bodhi so he doesn't have to start from scratch.

🐙 "You basically wrote a summary for your AI every night." Every night. Like a bedtime story. Except instead of "and they all lived happily ever after," it ends with "run bun run dev and check the Granola MCP access first."

2. /wrap — The Session Close Ritual

/wrap is a slash command I built into my PAI system. When I type it at the end of a session, Bodhi executes a full ritual automatically:

Update HANDOFF.md — dates, what we built, next priorities
Push all modified repos — nothing gets lost
Update Open Brain — project statuses + key thoughts embedded as vectors
Report completion — I get a ✓ checklist before closing

It took maybe 4 hours to build. It saves me approximately 20 minutes of context-setting per session. That's a 5x ROI on day one and it compounds forever.

The rule: never compact silently. Fresh session > degraded session.

If Bodhi is running out of context in the middle of something, he alerts me, updates the HANDOFF, pushes everything, and we start clean. No half-baked sessions pretending to function normally while the lights dim.

3. SessionStart Hook — The Opener

On the other end: the SessionStart hook runs automatically when a new Claude Code session begins. It reads HANDOFF.md and presents a brief:

📋 What We Did Last Session:
- Built the Granola deal intake feature
- Deployed 4 articles to gradyhodge.com
- Fixed the YAML colon bug (oops)

➡️ Next Up:
1. Lab articles (you're reading one right now)
2. TELOS path bug fix
3. Publish more articles

This means Bodhi starts every session already knowing where we are. No warm-up. No explanation. We just build.

What I Learned

The gap between "using AI" and "having an AI infrastructure" is this:

Does your AI remember you tomorrow?

If the answer is no, you're using a vending machine. A very smart vending machine, but still.

The /wrap protocol turns episodic conversations into a continuous relationship. That's not a small upgrade. That's the whole game.

The Dad Joke You Didn't Ask For

Why did the AI forget my name?

Because I didn't give it a memory foam mattress.

🥁 I'll show myself out.