The scenario below is an illustration, not a client account.
A bookkeeper tells her AI never to round client invoices to the nearest dollar. She told it last month. She told it the month before. Each time it agrees and complies, and each new session starts rounding again. Every month she spends the same five minutes teaching the same rule to something that has no way to keep it.
How should an AI system remember the corrections I give it?
It should keep a correction ledger: a durable, owned record where every fix you make is written as a scoped rule, together with the check that tells you when a draft has broken it. The workflow reads that record before the work ships. A draft that fails the check gets held instead of sent. Most AI setups have no such record, so nothing exists for the next draft to be measured against, and every session starts from zero.
Memory is the wrong word
The market sells “memory” as the fix for forgetful AI, and the word hides the real problem.
Memory, as it is sold, means the tool stores facts about you and pulls them back into conversation when they seem relevant. That has narrow uses. Storing a fact is a different job from keeping a standard. Your city and your job title are facts. “We never open a client email with a question” is a standard, and a standard has to be enforceable somewhere to mean anything.
A correction is closer to a transaction than to a memory. You teach the system something, and the lesson gets posted: written down, dated, and placed where the work it governs can be checked against it. A tool that resets and a system that learns differ on exactly this point.
What a correction ledger is
A correction ledger is a durable, owned record of every correction you have made to your AI. Each entry carries the scoped rule, the criterion a later draft can be checked against, and the condition under which the rule should be retired. The ledger is the mechanism behind “AI that learns from your corrections.”
The word “ledger” is deliberate. I spent 36 years building mission-critical bank systems, and the ledger is the object that discipline organises itself around. It is append-only. Every entry is dated, and nothing changes without a record of what changed and why.
That describes a design. It is not a claim about what your AI will do. Copying the shape of a bank ledger will not make a one-person workflow as reliable as a bank’s. What it does is make the workflow legible, and legibility has to come first.
How a correction becomes a ledger entry
Here is what should happen the moment you correct your AI.
- You correct the output. “We don’t talk to clients like that.” “That number is wrong.”
- The correction gets posted as a durable rule, stored outside the conversation, dated and owned. A line in a chat transcript you have to remember to re-read does not count.
- The rule enters the workflow’s working standards, the way a coding standard enters how a team writes software.
- Before the next draft ships, the workflow reads the ledger and checks the draft against the rule. A draft that breaks it gets held, with the failing entry named.
Step 4 is where almost every setup stops short. A pile of prompts does step one and drops the rest. Writing the rule down gives you a note. Giving the next draft something it can fail against gives you a control. A correction you keep is an investment; a correction you repeat is a tax.
Why the ledger has to be owned
The ledger only works if you own it, and this is the part people skip.
If your corrections live inside someone else’s tool, the ledger is theirs. Your judgment, written down one fix at a time, sits on their servers and disappears the day the subscription lapses or the company pivots. You end up renting your own standards back from a product. That is the renter’s trap.
An owned ledger lives inside your business, like software under version control. You can open it, edit a rule, retire a rule, and take the whole file with you when you change tools.
Your corrections are your judgment, written down
You did not get good at your business from a course. You got good from a few thousand small corrections, accumulated over years, until your standards were second nature.
Those corrections are your judgment, posted one entry at a time. On a tool with no ledger, none of it is written where the work can be measured against it, so a lesson you already taught has no way to reach the next draft and you teach it again. On a system with a ledger, the entries stay readable and yours. Whether they hold up is something you find out by running them against your own work over a few weeks. The ledger does not promise that in advance.
Try this now (4 minutes)
- Find a correction you have made to your AI more than twice.
- Write it as a one-sentence rule, dated today: “2026-08-25 / we always / we never ___.”
- Underneath, write the check. How would you know a draft had broken this rule? A rule you cannot check is a preference.
- Add the retirement condition. What would have to change for this rule to be wrong?
- Put all three lines in a file you own. That is the first page of your correction ledger.
Stop—this counts. You wrote the part most people leave out, which is the criterion a draft can fail. The workflow’s job is to read it before the work ships.
Which correction did you repeat this month? Reply and tell me the one you are tired of making. That is the entry your ledger should open with.
Frequently asked questions
Isn’t a chat history already a correction ledger? No. A chat history is a transcript you scroll back through, which gives today’s task nothing to be checked against. A ledger is written to be read by the workflow and turned into a check the next draft has to pass. One is a record you consult. The other is a standard the work gets measured against.
What if I post a correction wrong? Then you edit the entry or retire it. You can do that because the ledger is a file you own and audit. Nothing is buried inside a model where you cannot reach it. Reading a correction back and changing it is part of what engineering-grade buys you. The test is whether you can open the file and change the line yourself.
Is this the same as fine-tuning the AI model? No. Fine-tuning retrains a model, which is heavy, slow, opaque, and not yours to read. A correction ledger keeps a readable, owned set of rules that the workflow checks the work against at the moment it runs. You can audit it and you can edit it. The ledger is yours; the model’s weights are not.
Should a rule ever come out of the ledger? Yes, which is why each entry carries a retirement condition. A standard you set in March can be wrong by October because the client changed or the offer did. A ledger that can only grow will eventually enforce judgment you no longer hold. Write the exit condition at the same time you write the entry, while you still remember why you made it.