ASK TOBY

Your agents forget. Your files don't.

Your AI agent starts fresh every day.

That is not a flaw to fight. It is a setting you control.

Memory between sessions comes from one short file. It is how you make an AI agent remember.

The file carries the work forward. The agent reads it first and rewrites it last.

This guide shows what goes in these handoff notes, and what stays out.

1. Why a fresh session is a feature

A new session starts clean, with nothing left over from yesterday's detours.

You choose what carries over. That choice is the whole method of memory between sessions.

Decide it on purpose and every morning starts from the same good place.

2. One handoff file per agent

Give each agent its own file. One agent, one file.

Three headings are enough: open work, next step, blockers.

Open work is what is still in progress. Next step is the first thing to do on waking.

Blockers is what is waiting on someone else, and who.

That is all the agent needs to pick up where it stopped.

Here is a made up example of the handoff notes:

Open work: newsletter draft is written, the second section needs a fresh example.

Next step: finish section two, then send the draft for review.

Blockers: waiting on the owner to pick a subject line.

Three short lines. An agent that reads them knows exactly what to remember.

3. Keep it short on purpose

Set a size cap. A page or less.

A cap forces the useful lines to survive. Old detail drops out to make room for new.

A short file gets read in full every time. A long one gets skimmed.

If a line would not change what the agent does today, cut it.

4. Read first, rewrite last

Two habits make the file work.

Read it first, before anything else, every session.

Rewrite it last, before the session ends, so tomorrow's agent inherits today's state.

Write the file as a replacement, not a diary. The newest state wins.

Put both habits in your standing instructions so they never depend on a reminder.

5. Keep rules and memory apart

Standing rules go in one file. Today's state goes in another.

Rules change rarely: how to behave, what never to do, who decides what.

State changes daily: what is open, what is next.

When they share a file, the rules get buried and the state goes stale.

Two files, two jobs, both short.

6. Do not store what the agent can look up

Facts that live somewhere reliable stay there.

If a file, a record or a page holds the answer, point to it. Do not copy it into memory.

Memory is for what only the last session knew: where the work stands and what comes next.

Copies drift. Pointers stay true.

7. Check it

Before you trust the file, test it.

Start a fresh session and ask the agent what it will do next.

Compare the answer with the file. They should match.

If they do not, the file is unclear. Fix the file, not the agent.

Run this check whenever you change how the file is written.

Next

See your whole crew on one screen

Put your agents on a schedule

Get the free field guide and set up your first approval list.

Your Crew, Your Call field guide cover
FREE · PDF

YOUR CREW, YOUR CALL: THE FIELD GUIDE

How I run a fleet of AI agents from one screen and my phone.

One email with the guide. No spam, no list.

Want the full playbook? The Blueprint has the templates I use.

Questions people ask

Why do my agents start fresh every morning?
A short handoff note beats a long, muddy memory. Each agent starts the day clean and reads its own note. The note holds what's open, what's next and what's blocked. Short means it gets read. Muddy memory gets guessed at. Fresh start. Clear note.
How do my agents share memory?
Not with one magic brain. With a rule book and a record. Every agent loads the same rule book, and every decision I make is logged in one place. When an agent learns something worth keeping, it gets written into the rule book. Memory you can read beats memory you have to trust.

Jeff Sperandeo. Video editor. I ran the paid ads side and the organic side of a big YouTube channel at a Fortune 500 company. I saw how ads and AI find people from the inside. Now I build agents on my own machines, and I show you how.