This is the second post in a three-part series on how and why an AI coach needs a different kind of memory than a chatbot. Chatbots store facts about you; a coach needs to remember how you’ve grown, and that takes a different kind of memory. Read part 1, on why an AI coach needs a different kind of memory than a chatbot.
Memory has always been a tricky thing to factor into any artificial intelligence tool. Long before large language models came along, computer scientists and cognitive researchers had grappled with how to properly incorporate working memory and long-term memory into AI. So when we set out to design memory for our AI Coach, we knew we needed to integrate four types of memory:
- Working memory: staying present in the conversation
- Episodic memory: what happened, and what it meant
- Semantic memory: a person’s world and the people in it
- Procedural memory: learning how to coach this person
Working memory: Staying present in the conversation
Working memory describes the small, active workspace where you hold what you're thinking about right now. For a language model, the equivalent is the context window — the text the model can see when it writes its next response. The tempting approach is to fill the context window with everything the model might need. But that runs into two problems.
First, models don’t pick up details evenly as inputs get longer. Researchers showed in 2023 that model performance drops when the relevant information sits in the middle of a long context window rather than at the beginning or end, so more context can make a response worse. Second, every extra page of context the model has to read adds latency. That matters a great deal in a coaching conversation, where a long pause lands very differently than it does when you're waiting on a research report. BetterUp’s own research points the same way: in the CANDOR corpus, more than 1,600 unscripted video conversations between strangers, the speed of turn-taking was linked to how connected people felt afterward.
So instead of handing the coach an ever-growing transcript and hoping it notices what matters, our AI Coach keeps a structured understanding of where the conversation is, what the person is working on, and what has already been covered. It builds what the coach sees on each turn from that, and the coach spends its attention on what’s important instead of rereading and re-summarizing the entire conversation to respond to the user’s latest message.
Episodic memory: What happened, and what it meant
Episodic memory is your memory of experiences you've lived through, and it’s the heart of coaching memory. After each conversation, AI agents turn it into a structured record of the coaching that took place, and that record is much richer than a summary. It captures things like what the person was working through, what they were trying to achieve, what shifted for them as they talked, and what they decided to try next. Say a member spends a session preparing for a hard conversation with their manager. Weeks later, the coach remembers more than the topic. It remembers that they were worried about sounding defensive and planned to open with a question. The coach can recall what a moment meant to the person as well as what was said.
Episodes also arrive from outside the chat. When a performance review or 360 lands in the HR system, or the member has a session with their human BetterUp coach, that becomes part of the same record. So do conversations you’re having with your team on Teams/Slack or dialogue from meeting transcripts. These sources are critical since many BetterUp members work with both a human coach and our AI Coach. Because both draw on shared context, the AI Coach can pick up on Wednesday from what came up with their human coach on Monday. That's only possible when one system holds the whole coaching relationship.
Memory map
The memory behind our AI Coach
Four forms of memory, the sources that feed each one, and the two processes that keep them current and put them to use.
| Form of memory | In our AI Coach | Why it matters | What feeds it |
|---|---|---|---|
| Working memoryStaying present in the conversation | A structured understanding of where the conversation is, what the person is working on and what has been covered, rebuilt for the coach on each turn instead of an ever-growing transcript. | Models attend unevenly to long inputs, and every extra page adds time to a live conversation. | Coaching conversations |
| Episodic memoryWhat happened, and what it meant | A structured, multi-facet record of each conversation, plus episodes that arrive from outside the chat, like a new review, a 360 or a session with a human coach. | The coach can recall what a moment meant to the person as well as what was said. | Coaching conversations; sessions with a human BetterUp coach; BetterUp 360, reviews and the Whole Person Assessment; performance reviews and 360s from the HR system; Microsoft Teams and Slack conversations; meeting transcripts |
| Semantic memoryA person's world and the people in it | The people a member talks about and what they've shared about each, along with their role, goals and organizational context. | A member doesn't have to re-introduce their manager or recite their last review before the coach can help. | Coaching conversations; BetterUp 360, reviews and the Whole Person Assessment; performance reviews and 360s from the HR system; Microsoft Work IQ (meetings, deadlines and working relationships) |
| Procedural memoryLearning how to coach this person | A profile of each person that starts with onboarding and the Whole Person Assessment and grows with every interaction, shaping how the coach works with them. | The coach starts from a profile rather than a blank page, and keeps learning how to work with the person from there. | BetterUp 360, reviews and the Whole Person Assessment; onboarding; coaching conversations |
Between sessions
Consolidation
Agents work through each conversation, new data from integrations and human coach sessions in the background, connecting episodes and noticing patterns.
In the moment
Retrieval in time
The coach decides when context will help, asks its memory questions scoped in time, and synthesizes change. New insights become memory on the spot.
Each form of memory draws on established cognitive science: working memory (Baddeley & Hitch, 1974; Liu et al., 2023), episodic and semantic memory (Tulving, 1972) and procedural memory (Laird, Newell & Rosenbloom, 1987).
Semantic memory: A person’s world and the people in it
If episodic memory is what you’ve lived, semantic memory holds what you know. For a coach, this is some of the most important knowledge: a person’s role, their goals, their organization, and above all the people they work with. Coaching conversations are full of other people — a manager who keeps moving the goalposts or a direct report who's struggling. A coach who has to ask "who's Dana again?" loses the thread and could come off as being less attentive and trustworthy. That’s why we made it possible for the coach to remember the people a member talks about and add people to that growing list. Some of this knowledge can come from outside the conversation too, if an organization chooses to connect the AI Coach to other systems: past performance reviews and 360 feedback from Workday, results from BetterUp’s assessments, and context from Microsoft Work IQ, like a member’s upcoming meetings, deadlines and key working relationships. A member doesn't have to recite their last review or explain what’s happened that week before the coach can be useful.
Procedural memory: Learning how to coach this person
Procedural memory is know-how, or the skills you exercise without thinking about them like riding a bike or tying your shoes. The coaching equivalent is learning how to work with a particular person, such as whether they want candor or room to think, and how much support they need before a challenge will land. A good human coach learns how to read and understand someone over months of sessions. Our AI Coach doesn’t have to wait that long. It starts from a profile rather than a blank page, and from what BetterUp has learned across more than 5 million coaching sessions.
When a member takes BetterUp's Whole Person Assessment, the coach gets a picture of their strengths and growth areas across a range of behavioral dimensions. It uses that picture to understand the person, and also to shape how it coaches them. From there, the AI coach keeps refining how it works with that person as conversations accumulate.
After memory shipped
The median conversation doubled in length.
Together, these four types of memory let our AI Coach do what a good human coach does over months: remember what happened, know the people in your world, and learn how you like to be coached.
Up next: A look at how our AI coach chooses what to bring to the conversation and what to leave behind.
See how BetterUp’s AI Coach remembers each person’s growth, not just their facts. Talk to BetterUp