AI agent memory: what an assistant should remember about your business.
Useful memory is not a transcript of everything anyone ever typed. It is a short, scoped, dated record of what the business decided, where each fact came from, what is still owed, and what the assistant is not allowed to do. Here is the policy, a template to copy, and an illustrative example.
Good AI agent memory keeps decisions and sources, not every word.
An AI assistant working for a business should remember what saves people from explaining the same context twice: each project's goal, the standing decisions and who made them, stable preferences, approved facts with the file they came from, open commitments, receipts for completed actions, and the boundaries and approvals the owner has set. Every entry needs a scope, a date, a source, and a person who can change it.
It should never hold passwords, API tokens, card numbers, personal data it does not need, or other people's private content. A remembered fact is a pointer, not a permission: before quoting a price or sending a file, the assistant checks the live source, and anything public, financial or hard to undo still waits for a fresh yes. The NIST AI RMF Core lists outcomes such as documented roles and responsibilities, defined processes for human oversight, and privacy risk that is examined and documented. A written memory policy is a practical way to meet them for an assistant.
Decide what AI agent memory should hold for a business.
Start with the question the assistant would otherwise ask again next week. If the answer is stable, reusable and safe to keep, it belongs in memory. If it is a one-off detail, a guess, or something that changes by the hour, the assistant should read it from the live system instead.
- Scoped goals: what a project or client relationship is for, with the date the goal was set.
- Standing decisions and who made them: for example, which quotes go to the owner first, with the decider's name and the date.
- Preferences: tone, format, meeting hours, a contact's preferred channel, each marked with whose preference it is.
- Approved facts and their source files: prices, services and policies, each linked to the document that makes it true, such as the current price sheet.
- Open commitments: what was promised, to whom, by when, and what closes it.
- Receipts of completed actions: what was sent, booked or filed, with the identifier the other system returned.
- Boundaries and approvals: which actions may run alone, which always need a yes, and what the owner declined, each approval stored with its exact scope.
Raw chat logs are missing from that list on purpose. Tectomight's homepage describes timestamped raw history kept as the original source, with candidate preferences and decisions passing review before they become durable facts. The working memory the assistant reads before acting should be that curated layer, small enough for a person to read. The AI operator blueprint puts it from the engineering side: context windows are not memory, and decisions, preferences, client facts and past outcomes belong in files or a database the system reads at the start of every run.
List what an AI assistant must never store in memory.
Some information should never enter an assistant's memory, however convenient it looks. The test is simple: if a leaked entry would let someone log in, pay, impersonate a person or expose something private, it does not belong there.
- Passwords, one-time codes, recovery phrases and API tokens. Access should run through the provider's own sign-in and a connection that can be revoked.
- Card numbers, bank account numbers and security codes. Record where payment is managed, not the number.
- Personal data the job does not need, such as home addresses, dates of birth, health details and identity documents.
- Other people's private content: a third party's forwarded thread, a colleague's direct messages, a client's confidential files. Remember where the document lives, not what it says.
Major assistants draw similar lines. Claude's help page says some information is never saved to memory even if the user asks, including government ID numbers, criminal history, financial account numbers and immigration status, and that sensitive topics such as health are not stored by default. For personal data covered by the GDPR, Article 5 requires data to be adequate, relevant and limited to what is necessary for the purpose. Whatever law applies to your business, the principle transfers well to memory: keep the least that does the job. On access, Tectomight's homepage says apps connect through the provider's own sign-in, Tectomight never sees a password, the token lands only in the seat of the person who connected it, and access can be revoked from the provider at any time.
Scope AI assistant memory per business, project and person.
Memory without scope leaks. One client's preference turns up in another client's proposal, a decision made for one project gets applied to the next, or a private note surfaces in a group chat. The fix is to give every entry an address before it is saved.
Use three levels. Business memory holds what is true company-wide: approved services, standard terms, brand rules, and who can approve what. Project or client memory holds the goal, decisions, files and commitments for one piece of work, and is loaded only when that work is active. Person memory holds one individual's preferences and permissions and stays with that person unless the policy says otherwise. When an entry could fit two levels, save it at the narrower one.
Claude's help page says each project has its own separate memory space and project summary, kept apart from other projects and from chats outside projects. Tectomight's homepage describes a similar structure: each project gets its own page with the goal, the live plan, the decisions waiting on the owner, its files and its receipts, and the operator loads that project's context before it answers. On the Business plan, each person gets their own context and permissions over one shared company memory.
Date every memory, and decide when it expires.
A fact without a date cannot be trusted for long. Prices change, people change roles, and a decision made for the busy season stops applying in the quiet one. Record when each entry was created, where it came from, and when it should be reviewed.
Give categories different lifetimes. Standing decisions and brand rules last until someone changes them, with a scheduled review. Prices and availability should expire fast or simply point to the live sheet. Commitments end when they close, and the closing receipt replaces them. Preferences nobody has used for a while get reviewed. Expired does not have to mean deleted: marking an entry as superseded keeps a history of what the assistant believed and when.
GDPR Article 5 expresses the same idea for personal data: keep it accurate and, where necessary, up to date, erase or correct inaccurate data without delay, and keep identifiable data no longer than the purpose requires. Tectomight's homepage says its memory review marks outdated facts as superseded while preserving their history.
Verify remembered facts against the live source, and settle conflicts by rule.
Memory tells the assistant where to look and what it believed last time. It does not tell it what is true now. Before an action that depends on a remembered fact, the assistant should read the source again: the current price sheet before quoting, the live calendar before offering a slot, the CRM record before emailing a contact, the thread before following up. If the live source disagrees with memory, the live source wins and the memory is corrected.
Conflicts need a rule, not a guess. A correction from the owner beats an inference the assistant made. A newer approved document beats an older one. A signed agreement beats a chat message. When two sources of equal standing disagree, the assistant stops and asks, with both versions and their dates in front of the person who decides, then saves the resolution with that person's name so the conflict does not return next week.
No memory system recalls everything, and retrieval can surface the wrong entry. Tectomight's homepage says so plainly: memory supports the answer, it does not guarantee perfect recall. That is why the check against the live source belongs inside the workflow.
Let the owner see, correct and delete what the assistant remembers.
If nobody can read the memory, nobody can fix it. The owner, or someone the owner names, should be able to see what is stored, correct an entry and have it apply from the next task, delete one entry or a whole category, and pause memory for a sensitive conversation.
Consumer assistants already document these controls, which makes them a fair baseline for a business tool. Claude's help page describes reading memory by topic in its settings, editing or deleting a topic, pausing memory, resetting it permanently, and incognito chats that are not saved to memory. Google's Gemini privacy hub says saved instructions stay saved until the user deletes them, that saving can be stopped at any time, and that a Temporary Chat is available for information the user does not want Gemini to remember.
Ask any vendor where you can read what it remembers about your business, how you correct it, and what happens to that memory when you stop. For Tectomight, the AI executive assistant page says the environment and its stored memory are removed entirely on request after cancellation, and the small business page adds that the deletion can be confirmed in writing. The NIST AI RMF Core lists mechanisms for capturing user input, appeal and override among its post-deployment monitoring outcomes, and a correction path for memory is one of them.
Make sure AI agent memory never grants new authority.
The most dangerous memory entry is one that reads like permission. "The owner approved sending the proposal" was true for one proposal, to one recipient, on one day. If the assistant treats that sentence as a standing rule, the next proposal goes to the next client without anyone deciding. Memory can record that an approval happened. It must not create a new one.
Write the rule into the policy. Every send, payment, publication, deletion or account change needs an approval that matches the exact action now, or a standing rule the owner wrote down on purpose, with its scope. A remembered preference, a past yes, or a note that something is usually fine does not qualify. Declines deserve memory too, so the same question is not asked again in new words.
Tectomight's pages state this boundary directly. Sensitive, financial, public, destructive or account-changing actions wait for the owner's explicit yes, every time. Anything public, financial or hard to undo waits on the owner's board as one yes or no with the reason attached, and a decline is remembered. Boundaries are agreed during onboarding and can be tightened later with one message. The sales follow-up guide applies the same split between preparation, approval, sending and proof.
An AI assistant memory policy template.
Scope levels: [business, project or client, person] Default: [the narrowest level that works]
Keep: [goals, standing decisions with name and date, preferences with their owner, approved facts with source file, open commitments with their closing condition, receipts with identifiers, boundaries and approvals with exact scope]
Never store: [passwords, codes, tokens, card and bank numbers, identity documents, personal data the job does not need, other people's private content]
Every entry carries: [scope, created date, source, who confirmed it, review or expiry date, status: active, superseded or expired]
Lifetimes: [decisions: until changed, reviewed on a schedule] [prices: read live or expire fast] [commitments: closed by the receipt] [preferences: reviewed when unused]
Before acting: [re-read the live source for every price, date, recipient, file or status the action depends on]
Conflicts: [owner correction, then newer approved document, then older document, then assistant inference] Equal standing: [stop, ask, save the answer with the decider's name]
Owner controls: [who can view, correct, delete and pause memory, and how deletion is confirmed when the service ends]
Authority: [memory records approvals and never grants them; sends, payments, publications, deletions and account changes need a matching approval now or a written standing rule]
Review: [who reviews memory, how often, and where corrections are logged]
One price change, one discount request, and what memory is allowed to do.
This example is illustrative, not a customer record. A small design studio uses an assistant for quotes and client follow-up. Its business memory holds an approved fact: the price of the standard logo package, sourced to the price sheet in the shared drive, with a review date at the end of the quarter.
Midway through the quarter, the owner updates the price sheet. The next day a new lead asks for a logo quote. The assistant finds the remembered price, but reads the live sheet first as the policy requires, sees the new figure, uses it, and marks the old entry as superseded with the date and the sheet as the reason. The draft cites the sheet, not the memory.
The lead then asks for a discount. Memory holds a note that the owner once approved a discount for a returning client. The assistant does not reuse it, because a past approval for someone else is a record, not a rule. The request goes to the owner as one decision with the context attached. The owner declines, the decline is recorded with its date, and the quote goes out at the listed price once the owner approves the final draft. Nothing in memory gave the assistant more authority than it had before the conversation started.
AI agent memory FAQs.
What is AI agent memory?
It is the information an AI agent keeps between tasks so it does not start from zero each time: goals, decisions, preferences, approved facts, open commitments, receipts and boundaries. In a business, each entry should be scoped, dated, sourced and open to review.
What should an AI assistant remember about a business?
Scoped goals, standing decisions and who made them, preferences and whose they are, approved facts linked to their source files, open commitments, receipts of completed actions, and the boundaries and approvals the owner has set.
What should an AI assistant never store in memory?
Passwords, one-time codes, API tokens, card and bank account numbers, identity documents, personal data the task does not need, and other people's private content. Record where something lives, never the secret itself.
How long should an AI agent keep a memory?
Only as long as it stays accurate and useful for its purpose. Give each category a lifetime, date every entry, review standing decisions on a schedule, and mark replaced facts as superseded instead of silently overwriting them.
Can an AI assistant act on something it remembers without checking?
It should not. A remembered price, date, recipient or status must be checked against the live source before any action that depends on it. If the source disagrees, the source wins and the memory is corrected.
Does a remembered approval let the assistant act again later?
No. An approval covers the exact action that was approved. Sends, payments, publications, deletions and account changes need a matching approval now, or a standing rule the owner wrote down on purpose with its scope.
Should the business owner be able to see and delete AI memory?
Yes. Expect to view what is stored, correct an entry, delete one entry or everything, and pause memory for sensitive work. Consumer assistants such as Claude and Gemini already document controls like these.
Is AI assistant memory always accurate?
No. Retrieval can miss an entry or surface an outdated one, and facts change after they are saved. Dates, sources, regular review and a check against the live system before acting are what make memory safe to rely on.
The frameworks and controls behind the policy.
See what the assistant would remember, and what stays under your control.
Bring one workflow where context keeps getting lost or explained again. We will show you how Tectomight would handle it, what it needs access to, and what should still stay under your control.
For the service itself, see the AI executive assistant, AI agents for small business and the AI operator blueprint. For adjacent reading, see AI agent vs AI assistant and AI assistant for sales follow-up.
Book a live demo