An AI coach that forgets starts from zero every time. It asks the same opening questions, misses last week’s promise and gives advice that could fit anyone. An AI coach that remembers avoids all of that. But it can only do so because it keeps stored information about the people it coaches. Once a coach keeps that, the HR team that brought it in has something new to understand. The value of remembering and the exposure of keeping come from the same place, so you can’t weigh one without the other.
This post is the long version of one box in our Governance Canvas for AI coaching, the one on memory and retention. It belongs to a wider series on AI governance for people management. It covers five choices that every AI coach with a memory has already made. Each one is about what the coach keeps. HR doesn’t have to make these choices itself. It does need to understand them and judge them against what the coaching is for.
Why does an AI coach need to remember anything?
Coaching works over time. Say a manager tells her coach on Monday that she’ll stop rewriting her team’s reports. A good coach asks about it the next week. If the coach can’t remember the promise, it can’t follow up, and she is left to carry the thread alone. Memory also saves time. The coach doesn’t need to ask again about her role, her team or the problem she raised last month. It also lets the advice fit the person, because feedback for a new manager of four is different from feedback for a director of forty. A coaching plan that adjusts week by week depends on this kind of continuity.
There is some research behind the idea that the relationship matters. A 2020 meta-analysis by Graßmann, Schölmerich and Schermuly pooled 27 samples covering 3,563 coaching processes. It found that the working alliance between coach and client was linked to coaching outcomes at r = .41. The authors call that a moderate and consistent relationship. But it is a correlation, and it was measured with human coaches. It says nothing about memory and nothing about AI. Our own view is that being known and followed up is one plausible route to that kind of alliance. That is our inference, not a finding. We found no study that measures whether memory in an AI coach improves coaching outcomes.
One more limit needs saying plainly. Memory in most AI systems is not recall in the human sense. The system stores notes and re-reads them before it replies. So the notes are the thing to understand.
What is “memory” in an AI coach, exactly?
Most writing on this topic treats memory as one thing. It isn’t. An AI coach can keep several kinds of stored information, and each one does a different job. Each also costs something different to keep. The five layers below exist in some form across the industry, though not every coach keeps all of them. This is about what a coach keeps after a session. What it reads during one is a separate question, which our piece on AI coaching in Slack and Teams covers.
| Layer | What it is | What it makes possible | Who could plausibly reach it | What is at stake in keeping it |
|---|---|---|---|---|
| Raw transcript | Every word of the session, as said | The most accurate look back at what happened | The person; the employer, if the design shares it; vendor staff; the model provider while a reply is produced | The most candid and sensitive record, kept in full |
| Session summary | A short account of what was worked on, decided and left open | Follow-up from one session to the next, which is most of what continuity means | The person; the employer, often, where coaching is assigned; vendor staff | Someone or something chose what went in, and it is the layer most likely to be shared |
| Facts or topics | Short items pulled out about the person, such as a role, a goal or a habit | Advice that fits the person without a full record | The person; vendor staff; the model provider each time they are read into a reply | Facts go stale and can keep shaping advice after they stop being true |
| Scores and results | Assessment profiles, skill ratings and similar structured data | A fixed reference point for what to work on | The person; the employer, often; vendor staff | It looks like a judgment, so it is the easiest layer to reuse for another purpose |
| Aggregates | Patterns across many people, such as common topics in a team | A view of where a group needs help | The employer; vendor staff | Rarely names anyone, but small groups can make it traceable |
The raw transcript is the fullest record. It allows the most accurate look back. But coaching rarely needs it, because a coach seldom has to quote last month’s session to anyone. It is also the most sensitive layer. People tell a coach things they wouldn’t say in a meeting, about their boss or their doubts about the job.
Why the summary deserves the closest look
The summary matters most, because most coaching continuity is built on it. After a session, the coach writes a short account of what the person worked on, what they plan to try and what is still open. Next time, it reads that account before it says anything. That is what lets it ask how the talk with the difficult peer went.
A summary is far smaller than a transcript. But it isn’t neutral, because something chose what went into it. A summary can turn a passing remark into a stated plan. It can also leave out the context that made a comment fair. It is also the layer an employer is most likely to see, since a summary is the natural thing to share when coaching is part of a program. So the summary is where the coaching value and the exposure meet. A coach that keeps summaries and not transcripts hasn’t escaped the question. It has moved it to a shorter record that more people can read.
Facts or topics are the short items a coach pulls out and keeps. They might be a person’s role and team size, a goal, or a habit they want to change. They make the coaching feel personal without a full record. They can also go stale. A fact that was true in March may be wrong by June, and it can still shape the advice.
Scores and assessment results are the most structured layer. A personality profile or a skill rating gives the coach a fixed reference point. It also looks the most like a judgment about the person. That makes it the easiest layer to reuse for a purpose it wasn’t made for.
Aggregates sit across people, not inside one person’s record. They show patterns, such as which topics come up most in a team. They rarely name anyone. Even so, a team of five can make a pattern easy to trace back.
What public AI assistants show
Public AI assistants show that these layers exist in practice. They are general-purpose products, not coaching tools, so they illustrate structure only. Consumer and enterprise versions also differ. Anthropic’s help page on Claude’s memory describes memory saved as a set of individual topics. OpenAI describes ChatGPT’s memory as drawing on several sources. Two of them are memories it saves and reference to past chats. Google’s help page on Gemini’s memory says memory of past chats isn’t available on work, school or supervised accounts. None of this tells you how any coaching product works. It shows that “memory” in tools people already use is a set of layers with different rules.
Who can reach each layer?
Four parties could reach stored coaching data. The first is the person being coached. The second is their employer, through HR or an admin. The third is the vendor’s own staff, such as support or engineering. The fourth is the model provider, which is the company whose AI model writes the replies.
Most privacy language in this market is about the employer. It says what HR can and can’t see. That focus makes sense, because the employer question matters most for trust. But it often says little about the other two parties. The table shows that vendor staff can plausibly reach every layer. The model provider also sees whatever is sent to it when a reply is produced. Who on the vendor side can see what is a question your IT and security teams will have a view on.
The employer question has its own answer in our Governance Canvas, and this post doesn’t re-argue it. Self-directed coaching stays private to the individual. Assigned coaching and coaching linked to a decision about someone give HR summaries and engagement, not transcripts. The reasoning is set out in the Governance Canvas section on privacy tiers.
Why does this matter for the coaching itself? Coaching only works if people say what is really going on. We found no study on how open people are with an AI coach when they know their employer can read the transcripts. So the best we can do is reason from nearby work, and say that it’s nearby. Kalischko and Riedl (2024) ran a survey of 1,119 people about electronic monitoring at work. People who felt more monitored reported more privacy invasion, lower trust in their organization and more stress. That study is about monitoring work, not coaching, and it measures how people felt.
Amy Edmondson’s 1999 study of psychological safety points the same way at the team level. It looked at 51 work teams in a manufacturing company. Teams with more psychological safety, a shared belief that it is safe to take interpersonal risks, showed more learning behavior. The plain version is that people speak up where it feels safe to. Neither study is about AI, but together they support a simple logic: people tell a coach less when they think someone else is reading.
How long should a coaching memory live?
Every coach that remembers has a built-in answer to this, whether or not anyone wrote it down. There isn’t a correct number to copy. The better test is purpose. What is this layer kept for, and is it still needed for that? “Useful for the coaching” is a purpose. “Might be useful in a performance conversation later” is a different purpose. It needs its own case, and it changes what the coaching is.
The EU’s data protection law states the same test in public. Article 5 of the GDPR says personal data should be kept no longer than is needed for the purpose it was collected for. The law calls this storage limitation. The same article says data should be limited to what the purpose needs, which it calls data minimization. These rules cover personal data in the EU, and the UK has its own version. Read them here as the idea behind the choice, not as legal advice. Your counsel will have a view on how they apply to you.
Coaches tend to take one of a few broad approaches. Some keep their layers for as long as the coaching relationship is active. That gives the best continuity, but the record keeps growing for as long as the person uses the coach. Others keep layers for a set span and then let them go. That limits the record. But a person who comes back after a long gap may find the coach has lost the thread. Some keep very little beyond the session itself. That keeps the stored record small, and it hands the work of remembering back to the person. Each approach has a coaching cost. HR’s part is to judge whether the approach a coach takes fits what the coaching is meant to do.
What does deleting a coaching memory actually remove?
Deletion is less simple than it sounds in a layered memory, because the conversation and the memory taken from it are separate stored things. Deleting one doesn’t have to delete the other.
This can work in either direction, and both are fair designs. If deleting a conversation leaves the summary and the facts in place, the coaching carries on as before. The person has removed the record of what they said, but the coach still knows what it learned. If deleting a conversation also removes everything learned from it, the record is cleaner. But the coaching loses what it was built on. The coach may ask the same opening questions again, or miss a goal the person set weeks ago. Neither outcome is a trick. They are two ways of deciding what a deletion means, and the useful thing is to know which one a given deletion is.
Tools people already use show both sides of this. The same Claude help page says that deleting a conversation doesn’t remove the memories generated from it. OpenAI’s ChatGPT memory page says that deleting a chat alone doesn’t necessarily delete a saved memory created from it. Google’s Gemini Privacy Hub says chats that human reviewers have looked at are kept for up to three years, even when a person deletes their activity. That is a consumer product, and work accounts may have other terms. None of these is a coaching tool, and none of them is hiding anything. They are simply what layered memory looks like when a company writes it down.
Public documentation in general says little about aggregates or backups. That is an observation about how the topic tends to be written up, not a claim about any company. It does mean the full answer to “what does deletion remove” isn’t always on the page.
The law adds one more layer. Article 17 of the GDPR gives people a right to have their data erased. It is a right with stated grounds, and Article 17(3) lists exceptions to it. The article says nothing explicit about backups or aggregates. Workplace data is also often handled on legal grounds other than consent, which can change how the right applies. That is the general shape of the rule only, and your counsel will have a view.
Should memory ever train a model?
The fifth choice is about use. A stored layer can serve only the person it belongs to. Or it can also be used to improve the AI model behind the coach. Those are different things.
It helps to separate two kinds of contact with the model provider. When a coach sends a person’s words to a model to get a reply, the provider receives that data in transit. That is how most AI products work, and it’s why the provider sits on the map of who can reach each layer. Training is different. It means the data is used to change the model itself. What one person said could then shape what the model says to others.
Google’s Gemini Privacy Hub gives a public example. It says consumer activity may be used to improve Google’s AI, including training generative models. Those are consumer terms, and work accounts may be under different ones. Whatever a given product’s terms say, the useful point is that training is a separate choice from how long memory lasts. A coach could keep memory briefly and still train on it. It could also keep memory for years and never train on it.
Where does this get hard?
The benefit and the exposure can’t be pulled apart. A coach that keeps less gives shallower coaching. It asks more of the same questions and knows less about the person it’s talking to. A coach that keeps more can coach better, but there is more to look after. No coach gives full continuity with nothing stored. So every coach sits somewhere on that line, and HR’s job is to judge whether that point fits the coaching’s purpose. A short program for new managers and a long executive coaching relationship may fairly sit in different places.
Regulation makes this harder, not easier. The EU AI Act lists employment as one of its areas of concern in Annex III, under point 4. The listed uses include AI systems that make decisions about promotion or termination. They also include systems that assign tasks based on behavior or personal traits, and systems that monitor or evaluate performance. Article 6(3) adds that an Annex III system is always treated as high-risk where it profiles people. Coaching isn’t named in that list. But coaching memory could become a candidate for that category if what it holds starts to feed decisions about the employee. That is one reason some organizations give for keeping coaching data apart from performance decisions. Whether a given system falls into the category is a question for counsel, not for this post.
The honest answer for many organizations is that it depends. Works councils, local employment law and an organization’s own policies can all shift where the line should sit. An answer that fits a company in Texas may not fit one in Germany. I don’t think there is a universal rule here, and I’d be wary of anyone who offers one.
A framework can at least give the conversation a shape. The NIST AI Risk Management Framework is voluntary. Its core groups the work into four functions called Govern, Map, Measure and Manage. A team can use it to structure the discussion, though it sets no rules on how long anything is kept.
How Risely approaches this
Memory is part of how Merlin coaches, so what it keeps shapes the coaching as much as the privacy. Risely treats those choices as design decisions made with the purpose of the coaching in mind. We think about them in the same terms this post has used. Which layers are kept, who can reach them and are they still needed for the coaching they serve? We’d rather explain those choices than ask anyone to take them on trust. The details of how Merlin handles coaching data are set out in the Risely Trust Center.