Loading articles…
Field notes for building a business that can move without you.
We write from the intersection of operating systems, software, sales, marketing, and leadership—for owners building a more capable company.
Announcement
Memory belongs to a person. Knowledge belongs to the business.
Introducing Memory in Telos. Sapience now learns how each person works and writes, privately and visibly. What it learns about your business goes where your whole team can use it.
- Author
Alexander Ford- Published
- Reading time
- 8 min read
Contents
ChatGPT and Claude have become remarkably good at remembering you. Mention a project once, and it comes up naturally weeks later. Say you prefer short answers, and you get short answers. Both now let you see and delete what they remember.[1][2] For a personal assistant, that is close to the right design.
A business is a different problem. Listen to what someone tells their assistant over a working day. Some of it is about them: how they write, when they do their best work, what they're focused on this quarter. Much of it is about the business: a client's renewal date, the discount the owner approved, the new way you handle refunds, the reason a project slipped.
In a personal assistant, both kinds land in the same place: a private memory attached to one person's account. That is the right home for the first kind. It is the wrong home for the second.
Where business knowledge goes to hide
Much of the AI your team uses at work isn't AI the business chose. Microsoft and LinkedIn found that 78% of people who use AI at work bring their own tools, and 80% at small and medium-sized companies.[3] Each of those accounts is building its own picture of your business, one conversation at a time.
The privacy of that memory is deliberate, and it is right for personal details. Claude's documentation says workspace owners "can't view or edit a user's individual memories."[2] But the same privacy means that when one person's assistant learns your refund policy changed, no one else's assistant learns it, no one can check that it understood correctly, and the knowledge sits with that person rather than with the business.
This is the oldest knowledge problem in business, now automated. In a survey of more than 1,000 workers, Panopto found that 42% of institutional knowledge is unique to the individual, and that people lose 5.3 hours a week waiting for information from colleagues or recreating knowledge that already exists.[4] An assistant that learns a great deal but keeps it in one person's private memory feels like knowledge is being captured. In practice, it is being scattered.
There is a quieter problem too: drift. Your records say the price went up in September. An assistant's memory, formed in August, still holds the old one. Now there are two versions of the truth, and one of them is invisible.
Two kinds of knowledge, two homes
The question we ended up asking wasn't "how much should the assistant remember?" It was "where does each thing it learns belong?"
- Knowledge about a person is how they work, how they write, what they prefer and what they're focused on. It belongs to that person. It should be private to them, visible to them and theirs to correct.
- Knowledge about the business is clients, decisions, processes and commitments. It belongs in the system the whole team works from: the knowledge base, the CRM and the projects. There it is shared, it can be corrected once for everyone, and it stays when people move on.
A useful assistant draws on both. A trustworthy one never confuses them. Keeping them apart also makes the assistant better at its job. The engineers who build these models describe an AI's working context as "a finite resource with diminishing marginal returns."[5] An assistant does more with the right few facts, from the right place, than with everything it has ever heard.
How we arrived here
We didn't start with that line drawn. We reached it in three steps.
A notebook for each workspace. Sapience's first memory was a notebook the assistant wrote as it worked, shared across the workspace. Most assistants still work this way. This fall, we went through all 231 active notes across 26 client workspaces and asked of each one where it really belonged:
- 139 were working notes with no lasting value.
- 39 were about one particular person.
- 53, more than the personal notes, were business knowledge: processes, decisions and client context that belonged in the knowledge base or on a record, where the team could see and use it.
The notebook had become a second, unofficial system of record that nobody was maintaining. It held more knowledge about the business than about the people.
A learning system designed for every case. Our first design for learning about each person was thorough to a fault, with staged drafts, confirmation steps and readiness checks. It was built to handle every edge case before it had proven it could do the simple one well, and it never reached clients. We replaced it with a design simple enough to see and undo: one review each night, applied right away, with every change visible.
Learning that stays on its subject. Early runs of the new design surfaced the same line from the other side. Without a clear boundary, learning about a person picked up business events and other people's details along the way. So we made the boundary part of how learning works: personal memory learns only about the person. What is true about the business goes to the business.
Then we retired the notebook and moved each note to its home. Personal notes went to the person they described. Business knowledge went into each workspace's knowledge base and records. The working notes were archived.
What Sapience remembers now
About you, privately
Each night, Sapience reviews what's new: the messages you wrote in your conversations, the email you sent, and the edits you made to its drafts before they went out. The first night, it looks back 30 days. It learns from your words, not its own replies, and it skips conversations you've deleted.
From that, it keeps a short profile, covering how you work, how you communicate, how best to help you and how you write, along with a set of short memories, such as a preference or what you're focused on right now.
Everything is on one page, under Preferences → AI → Memory. You can edit your profile in your own words, add or remove memories, and undo a whole night of learning if it got something wrong. Tell Sapience to remember or forget something and it happens on the spot. What you tell it yourself, and anything you edit, is never overwritten by learning. Learning is on by default, and one switch turns it off.
Your memory is yours alone. Teammates can't see it, and their assistant doesn't use it. On a teammate's profile, people see only working relationships, such as who reports to whom.
About your business, in your workspace
What Sapience knows about your clients, decisions and processes comes from your workspace: your knowledge base, your CRM and your projects. When you tell it something the whole team should know, ask it to add it to the right record. From then on, every teammate's Sapience can find it, and a correction made once is made for everyone.
On infrastructure we run
The nightly learning runs on AI models we host on our own servers, so your conversations and email aren't sent to an outside AI company to build your profile. Each company's memory is separated from every other company's by the database itself. Learning doesn't use your credits.
The step-by-step guide is Manage what Sapience remembers about you.
What it means for your business
- Each person gets help the way they like it, without repeating themselves.
- Business knowledge builds up in one place your whole team uses, instead of scattering across private accounts.
- There is one version of the truth. Change a price or a process once, and every assistant in your workspace works from the change.
- What your business learns stays with your business when people move on.
What comes next
Learning happens overnight. What you tell Sapience directly applies immediately; what it learns on its own appears the next morning.
Today it learns from your conversations, your sent email and your edits to its drafts. We are deliberately slow to add other sources, such as meetings, because each one has to earn its place on a page you can see and correct.
We are also measuring whether memory makes Sapience more useful: how often people correct or remove what it learns, and whether its drafts need fewer edits. If it isn't helping, we'll say so and change course.
Why we build it this way
Your business's knowledge is one of its most valuable assets. It should build up in a system you own and your team can see, not in private accounts on someone else's platform. It's the same reason we keep your records on servers we own and operate, in our own building.
Software can learn how each of your people works. What it learns about your business should belong to your business.
If you're a Telos client, Memory is already on. Open Preferences → AI → Memory tomorrow morning to see what Sapience has learned about you. If you're not yet working with us and want a system built this way, see what Accelerator OS includes.
[1] OpenAI, "Memory in ChatGPT," OpenAI Help Center. help.openai.com
[2] Anthropic, "How does Claude's memory work?," Claude Help Center. On Team and Enterprise plans, memories remain exportable by administrators under the organization's data-retention policy. support.claude.com
[3] Microsoft and LinkedIn, "AI at Work Is Here. Now Comes the Hard Part," 2024 Work Trend Index, May 2024. microsoft.com
[4] Panopto, "Workplace Knowledge and Productivity Report," July 2018, survey of more than 1,000 U.S. workers. Archived release
[5] Anthropic, "Effective context engineering for AI agents," September 2025. anthropic.com
