All Things AI · Season 2 · The Models · Episode 8
Memory is not a model either. It is the layer that lets any model hold your rules, your corrections, and your past decisions across months, and consult the relevant ones automatically. It is the only part of this that gets better while you are not working on it.
Why this sits in the models season
Like the adapter, memory is not something you open. It is a capability you attach to whatever you already use, and it changes the behaviour of all of it.
The pairing is worth stating clearly because people conflate them. An adapter carries how you write. Memory carries what you have decided. One is style and the other is judgment, and having only one of them produces output that is either well-written and wrong or correct and characterless.
The three things called memory
They are not equally useful and the difference determines whether this works at scale.
Session history
What was said in the current conversation. Automatic, useful, and gone when you close it. This is what most people mean by the model remembers and it is the weakest form by a wide margin.
A standing instruction file
A document read at the start of every request. It survives, which is a real improvement, and it has a hard ceiling: everything in it is read every time, so it becomes expensive and starts crowding out the actual work once it grows.
Retrieved memory
Many small notes stored separately, searched at the moment of the request, with only the few relevant ones pulled in. This is the version that scales to years of accumulation without the cost per request rising.
How retrieval works, briefly
Each note is converted into a representation of its meaning rather than its words. Your request is converted the same way. The system finds the notes closest in meaning and hands those over.
The practical consequence is that a rule you wrote as never phrase it that way in written material will surface when you ask for something written, even though you used none of the same words.
That is the property that makes it durable. Keyword systems require you to remember the phrasing you used. Meaning-based retrieval does not, which is why it still works when the collection is large and old.
What belongs in it
The instinct is facts. The value is decisions.
- Corrections, with the reason attached. The single highest-value category and the one that makes the system compound.
- Standing rules about what is never acceptable, written once rather than remembered.
- Preferences with their history, because the reason is what lets a rule apply to a situation you did not anticipate.
- Context about ongoing work that exists nowhere else, which is most of the context that actually matters.
What to try in the first week
Two exercises, and the second one is the habit that makes this real.
Write twenty notes in one sitting from things you already know. The rules you would tell a new assistant on their first day. The things you always want, the things you never want, the way you like work presented. Twenty is enough to feel the difference immediately and it takes half an hour.
Then, for one week, every single time you correct something, write the correction down before you move on. Not later. At the moment. That habit is the entire discipline of this capability and it is the reason most people’s memory layer stays thin, because the moment of correction is exactly when you are least inclined to stop and record anything.
Keeping it consistent across machines
Once memory is valuable, having two versions of it becomes a real problem. You correct something on the laptop and the desktop carries on being wrong, which is worse than not having it, because you believe the correction is in place.
Two machines can reconcile on a schedule. The design rule that matters is union only: copy across, take the newer version, never delete. A sync with delete permission will eventually mirror an incomplete folder onto a complete one, silently.
Run it hourly rather than nightly if either machine sleeps, because a nightly job aimed at a sleeping laptop simply never runs, and you find out months later.
Where this goes wrong
Four failures, and three of them are habits.
Storing facts. They go stale, your files hold them better, and a stale fact retrieved confidently is worse than no memory at all.
Never writing the note. The system only knows what you tell it, and this is where the whole thing usually dies.
A sync that can delete. Takes a dated copy before every run or it will eventually cost you months of accumulation.
Treating it as authority. It records what was true when written. If a note names a tool or a file, that may be gone by the time it surfaces, so it is a strong prior rather than a fact.
How to write a note that is still useful in a year
Most notes fail not because they were never written but because they were written in a way that is useless when retrieved out of context.
The failure looks like this: do not do that. Perfectly clear at the time and meaningless three months later, because the thing being referred to is gone. Retrieval brings back the note without the conversation around it, so the note has to stand alone.
The shape that works is the situation, the rule, and the reason. When drafting anything that goes to a client, do not open with a question, because it reads as a survey rather than as a person. Anyone reading that in a year knows what it applies to, what to do, and enough about why to handle a case it does not exactly cover.
The reason is the part people skip and it is the part that does the work. A rule without a reason applies only to situations identical to the one that produced it. A rule with a reason generalises.
What to do when a note turns out to be wrong
This happens and how you handle it determines whether the collection stays trustworthy.
The instinct is to leave it, because deleting feels like losing something. That is how a memory layer slowly fills with contradictions, and a system that retrieves two rules that disagree is worse than one that retrieves neither.
Correct it in place and say what changed. The note becomes the current rule with a line noting the previous position and why it moved. That way you get the current answer and the history, and you do not rediscover the old reasoning six months later and change it back.
Reviewing the collection occasionally is worth the hour. Not all of it, and not on a schedule, but when something surprises you it is usually worth looking at what else you wrote around the same time.
Why this is the capability that compounds
Everything else on this site is as good on day four hundred as on day one. A model does not learn from your corrections. A local model is a fixed file. An adapter is trained once and then static.
Memory is the only part that improves while you are not working on it, because every correction is permanent and every permanent correction removes a class of future error.
That makes it the highest-leverage thing to start early, and the thing most agents start last because it is the least visible. There is nothing impressive to show anyone. There is just a system that is quietly better every month.
What it costs
Nothing if the search runs locally, which is the normal arrangement. Storage is a folder of small text files and the search runs on your own machine. The only real cost is the discipline of writing the note, and that cost is the whole product.
Questions agents actually type
Can AI remember my preferences between sessions?
Not by default. Session history disappears when you close the conversation. Persistent memory is a separate layer that stores notes and retrieves the relevant ones at the moment of each request, which is what makes it survive across months.
What is the difference between AI memory and a prompt file?
A prompt file is read in full every time, so it has a hard size ceiling and gets expensive as it grows. Retrieved memory stores notes separately and pulls in only the few that match the current request, so it stays cheap however much accumulates.
What should I put in AI memory?
Decisions and corrections with the reason attached, rather than facts. Facts go stale and your files hold them better. A reason is what lets a rule apply correctly to a situation you never anticipated.
How do I keep AI memory the same on two computers?
Reconcile them on a schedule with a union-only rule: copy across, take the newer version, never delete. Take a dated copy before every run, and run it hourly rather than nightly if either machine sleeps.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.