- What does your agent actually remember?
- Why is a wrong note worse than a forgotten one?
- Who owns the memory in your office?
- What does a checked memory look like week to week?
- What should never go into agent memory?
- The 1-hour setup that makes memory safe to grow
Your agent kept a note about your best client last month, and the note is wrong. The contact left the firm in September. The budget figure came from a draft the client never approved. The agent still opens each new task with that note at the top of its mind, and it will keep opening with it until a person reads the memory and fixes the page. Memory size is the setting vendors talk about. Memory accuracy is the job nobody in your office owns yet.
The direct answer: agent memory is a record your firm keeps, and it needs an editor. Give 1 named person the memory to own. Let that person read it on a set rhythm, correct the wrong notes, and hold the delete key for the rest. A small memory a person has checked beats a large memory nobody has read, each time it loads into a task.
What does your agent actually remember?
An agent with memory carries notes across tasks. It stores the facts it picks up about your clients, your projects, your prices and your habits, and it loads those notes into the next piece of work. The tools Canadian firms already run ship this ability today, each with its own shape. This magazine read 1 version closely in Notion agent Memories are private to each user. Decide what they may keep, where the memories sit private to each user and a person decides what the agent may keep. Other tools share memory across a team, or hold it inside a single project. The shape varies. The management problem stays the same shape in each tool. Notes gather on their own, at machine speed, during ordinary work. Nobody writes them on purpose. Nobody reads them on a schedule. They sit in the background of each task, treated as ground truth by the agent that saved them.
Why is a wrong note worse than a forgotten one?
A forgotten fact costs you a repeated question. The agent asks again, a person answers again, and the work moves on with a small delay. A wrong note costs more, and it costs in silence. The agent works from the wrong client name, the old price or the dead deadline with full confidence, and the error rides inside finished work where a reviewer has to hunt for it. Memory errors also compound. A note saved in week 1 shapes the draft in week 2, and the week 2 output becomes the source for the note saved in week 3. By the end of a month, a single bad note can sit under a pile of work that looks consistent, because each layer agreed with the layer under it. Agreement feels like proof when a person reviews the pile. It is agreement with a mistake. The fix is a person reading the memory itself, at the source, before the pile grows. That reading habit is the same discipline this magazine set for finished work in Pick one project marked “done” and ask for the record it should have left behind. Ask for the record. Then read the record.
Who owns the memory in your office?
One named person, with the job written down. The owner holds 3 duties. The owner reads the memory store on a set rhythm. The owner corrects or deletes notes that fail a check against the source. The owner decides what classes of fact the agent may store at all. Pick the owner by proximity to the work the agent does most. A sales agent’s memory belongs with the person who runs the pipeline meeting. A support agent’s memory belongs with the person who reads the escalations. The owner needs 1 standing hour a month and the authority to delete. A shared duty fails here the way shared duties fail on the floor. When each person assumes a colleague read the notes, nobody reads the notes. The checker role has a cousin in this magazine’s earlier work on review, in The AI draft lands in seconds. Then someone senior spends the afternoon fixing it. That piece named a checker for output. This piece names an editor for the input the output grows from. The 2 roles pair well in 1 person for a small team.
What does a checked memory look like week to week?
A checked memory has a rhythm a new hire could follow. Each week, the owner scans new notes for the 3 failures that matter. A note names a person who left, a figure that changed, or a promise the firm never made. Each month, the owner reads the full store and cuts it hard. Notes with no source go. Notes older than the work they describe go. Notes that restate a live system of record go, because the agent can read the live record at task time and a stored copy can only drift from it. Each quarter, the owner reviews the classes. Which kinds of facts earned their keep, and which kinds caused corrections downstream? The classes that caused corrections lose their storage rights. The store shrinks as the rules sharpen, and that shrinking is the sign the system works. Teams picking the frame an agent runs inside can pair this rhythm with the control questions in Six frameworks want to run your agents. Pick the one your team can run half-asleep. A frame your team can run half-asleep, holding a memory your editor can read in an hour, is a system that survives a busy season.
What should never go into agent memory?
Draw the line before the agent draws it for you. Keep personal data about clients and staff out of stored memory unless a task truly needs it, and hold what stays under the same care your firm owes customer records under Canadian privacy law. A memory store is a record. Treat it as one a regulator, a client or a court could one day ask to see. Keep credentials, access codes and payment details out in full, with no exceptions for convenience. Keep guesses out. An agent that stores its own inferences as facts builds the compounding problem from the section above. Where the tool allows it, store facts with their source and date attached, and let inferences expire fast. A note that reads as a fact and began as a guess is the most dangerous line in the store.
The 1-hour setup that makes memory safe to grow
Set aside 1 hour this week and do 5 things in order. Open the memory store in the tool your team uses most, and read every note in it. Delete each note you cannot trace to a source you trust. Write the 3 classes of fact the agent may store, in plain words, where the team can see them. Name the owner, and put the monthly read on that person’s calendar as a standing appointment. Set the store as small as the tool allows, and grow the limit only after 2 clean monthly reads. That hour converts memory from a background habit into a managed record. The agent keeps its usefulness, the notes keep a person between them and your clients, and the next wrong note meets an editor before it meets a task.
The recap fits on an index card. Memory grows on its own. Accuracy grows only where a person works at it. An editor, a rhythm and a delete key turn a pile of notes into a record your firm can stand behind.
Name your memory editor this week, before the store grows past the size 1 person can read.
