I Taught My AI to Forget (Why Memory Needs a Garbage Collector)

TLDR

Everyone building with AI is racing to give it more memory. I did too. My AI setup now holds about 700 memories: preferences, corrections, project history, lessons. And the most useful thing I built for it this year wasn’t more memory. It was forgetting. Every memory now carries the date it was last verified, an expiration date based on what kind of fact it is, where it came from, and how confident we are in it. A compiled page decides what gets loaded each day under a hard size budget, weighted toward whatever I’ve actually been working on. On the day I checked, about one in nine memories made it into the room. Nothing is deleted. Old facts just stop speaking unless they’re asked. For an AI that’s going to work with you for years, I think forgetting is the frontier nobody’s talking about.

The Memory That Wouldn’t Stop Growing

Back in February I wrote about what your AI actually needs from you, and one line from that post has aged better than anything else I’ve written: “The soul doc is a constitution. Memory is case law. Lessons and wins are precedent. Over time, the case law and precedent do more work than the constitution does.”

It turned out to be true. It also turned out to be a problem.

Case law piles up. Every correction, every preference, every “remember this for next time” became a memory, and every memory wanted to be loaded into every session. The page that loads at the start of each session kept blowing past its size limit. On September 1, after trimming it three times in one day, I said the obvious thing: “we keep having to trim.”

The page was being asked to be three different things at once: who I am and how I work, what’s happening on every project right now, and an index of everything else we’ve ever learned. Those three things have completely different lifespans. My name doesn’t expire. Last Tuesday’s project status expires fast.

A Wrong Memory Is Worse Than No Memory

Here’s the part that made me take this seriously.

A missing memory costs you a question. The AI doesn’t know something, so it asks, or it looks it up. A wrong memory costs you a confident mistake. The AI “knows” something that stopped being true, and it acts on it without checking.

I have a real example. For months, the logs in my tarot app recorded which AI model answered each reading. Based on those logs, a conclusion got written down: the backup model had never been needed. Then we discovered the label in those logs was hardcoded. It said the same model name no matter which model actually answered. So the conclusion was based on a constant, not a measurement.

The fix to the code was simple. The fix to the memory was the interesting part. A new memory now exists specifically to overrule the old conclusion: anything the old logs said about backups, before September 5, proves nothing. Without that correction, a future session could have found the old conclusion, trusted it, and made a decision on it.

That’s the real risk of a long-lived AI: not that it forgets, but that it remembers something stale with total confidence.

Every Memory Gets a Birth Date and an Expiration Date

So every memory now carries a small header of facts about itself:

• When it was last verified. Not when it was written. When someone last confirmed it was still true.
• How long it’s good for, based on what kind of memory it is. A preference about how I like reports written lasts a long time. A note about a project’s current status doesn’t.
• Where it came from. Did I say it directly, was it observed in the work, or was it imported from somewhere else?
• How confident we are. High, medium, or low.
• Whether it’s still active, or whether a newer memory has replaced it.

When a newer memory supersedes an older one, the older one is retired automatically the next time the page compiles. When a memory passes its expiration date, it drops off the daily page. A review tool can then surface expired memories so a person can decide: verify it and extend it, update it, or let it go.

There’s also a guard that runs every time a new memory gets written. It warns if the memory is missing those fields, or if it looks like a near-duplicate of one that already exists. A lot of memory bloat is the same lesson written down four times.

Attention Decides What’s in the Room

The daily memory page is compiled now, not hand-edited. A script builds it from every memory’s own header, under a hard budget of 23,000 bytes.

A few memories are pinned and always make the page: who I am, the hard rules, the things that must never be forgotten. Some are marked to never appear on the page. Everything else competes for the remaining space based on how recent it is, what kind of memory it is, and whether someone marked it as important.

Then there’s attention. Space for each project follows where my work has actually been over the last two weeks, with a seven-day half-life. The project I’ve been working on all week gets more room. The one I haven’t touched in a month shrinks toward a minimum. And when I open a session inside a specific project, that project’s newest memories get brought in automatically.

On the day I checked, 82 of 713 memories were on the page. The other 631 weren’t gone. They just weren’t in the room.

It’s closer to how people work than any “remember everything” system. You don’t walk into a meeting thinking about everything you’ve ever known. You walk in thinking about the last few weeks, plus the handful of things that are always true about you.

Forgetting Isn’t Deleting

This is the key design choice, and the thing I’d tell anyone building memory for an AI: forgetting should mean demoting, not destroying.

Nothing that falls off the daily page is deleted. It moves to a full index that can be searched. And every time I send a message, a recall step checks my words against the whole memory set and pulls in anything relevant, even if it didn’t make the daily page. So an old lesson about a specific client or tool shows up when I mention that client or tool, and stays out of the way the rest of the time.

On top of that, any old fact gets treated with suspicion before it’s acted on. The instruction my AI works under is plain: recalled memories “reflect what was true when written,” and if one names a file, a function, or a setting, check that it still exists before recommending it. Decisions that have been waiting on me for more than seven days get re-checked against reality before they’re shown to me again.

The shape of it: remember everything, load only what matters today, and verify anything old before trusting it.

Why This Is the Next Frontier

What follows is my read on where this goes, not established fact.

Right now, most AI memory features are built around one question: how do we get the AI to remember more? That was the right question when AI forgot everything between conversations. It’s the wrong question for an AI that’s going to work alongside you for years.

Over a long enough time, every memory system faces the same three problems:

1. Staleness. Facts that were true when they were written stop being true, and nothing tells the AI.
2. Crowding. Old, low-value memories take up the space that current, high-value ones need.
3. Confidence without provenance. The AI can’t tell a fact I stated directly from one it guessed six months ago.

Every one of those gets worse with more memory, not better. The fix isn’t a bigger memory. It’s a memory with a lifecycle: birth dates, expiration dates, sources, confidence, retirement, and a budget for what’s in the room.

My bet is that within a few years, the best AI assistants will be judged less on how much they remember and more on how well they forget. The ones that know what they know, how they know it, and when they last checked will be the ones people trust with real work.

To stretch my own metaphor from February: case law needs a court of appeals. Some precedents get overturned. A legal system that could only ever add rulings, and never retire one, would eventually contradict itself into uselessness. So would an AI.

How to Teach Your Own AI to Forget

You don’t need my setup. You need the principle. If you keep any kind of memory or notes file for an AI you work with, try this:

1. Date every fact with when it was last confirmed, not just when it was written.
2. Give each kind of fact a shelf life. Preferences last. Project status expires in days or weeks.
3. Mark where each fact came from, and how sure you are.
4. When something changes, write the replacement and retire the old one. Don’t leave both around to argue.
5. Put a budget on what loads by default. A small, current page beats a giant, stale one.
6. Keep the rest searchable. Forgetting from the daily page shouldn’t mean losing it.
7. Re-check old facts before acting on them. Especially file names, settings, and anything that changes.

Your AI will get smarter the day it stops trying to remember everything at once.

I appreciate you reading this one. It’s the weirdest post in the series and maybe my favorite. If you want an AI assistant for your business that stays accurate for the long haul, not just impressive in the first week, book a call from the home page.