This is the first 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.
In the movie 50 First Dates, the main character Lucy wakes up every morning believing it's the same Sunday in October. A car accident left her unable to form new memories, so everything from the day before, including the man falling in love with her, is gone by the time she opens her eyes. Eventually Henry (the man falling in love with her) makes a videotape she watches each morning to catch up on the life she's been living, and it works, sort of.
I've been using this Lucy analogy to explain AI memory because the way she lives each day — still entirely herself but starting over every morning — is close to how a large language model meets each new conversation. The model arrives with enormous general knowledge and holds a conversation fluently, and then the conversation ends and none of it carries forward. Most of what the industry calls memory is some version of Henry's tape: a store of facts about you, written down along the way, and played back at the start of the next conversation. For a general-purpose assistant that's a reasonable design.
Coaching needs something else entirely.
What’s the difference between semantic and episodic memory?
Psychologists have been untangling different ways we remember since neuroscientist Endel Tulving drew a line in 1972 between semantic memory (the facts you know) and episodic memory (the experiences you've lived through). Coaching runs on both, but semantic memory is insufficient in itself because it’s essentially a fact store. Episodic memory is where the magic happens based on what a person makes of those episodes over time, which is much harder to build.
When someone works with a good coach, much of the value comes from the coach holding the thread of their development between conversations: the situation they were dreading, what they decided to try, how it went, what it revealed, and how they've changed since. Imagine a first-time manager who three weeks ago was preparing to tell a senior engineer his promotion case wasn't ready. She rehearsed the conversation twice and decided to open with what he'd done well. Last week she mentioned, almost in passing, that he's been quiet in standups since. Semantic memory would remember she manages a team of six and likes concrete advice. A coach remembers the arc and connects the quiet standups to the earlier promotion conversation. The coach understands the manager is working on how to deliver hard news without losing people, which is a question about who she's becoming as a leader — and bigger than any one conversation.
Illustrative example
Same conversations, two kinds of memory
What our AI coach keeps
- Three weeks agoPreparing to tell a senior engineer his promotion case isn't ready. Rehearsed it twice and chose to open with what he'd done well.
- Last weekMentioned, almost in passing, that he's been quiet in standups since.
- The threadLearning to deliver hard news without losing people.
What a chatbot memory keeps
- Manages a team of six
- Engineering org
- Prefers concrete advice
- Has a senior engineer on her team
Accurate, and useful for smoothing the next request. None of it connects the quiet standups to the conversation three weeks earlier.
That kind of memory has to capture both what happened and what it meant, it has to reason about how things change over time, and it has to know when bringing something back will help and when it will only get in the way. It also has to draw on more than the chat window, because a lot of what shapes someone's growth at work, like their last performance review, their 360 feedback, or what's actually on their plate this week, never gets typed into a conversation.
A difference in behavior
We built our AI coach's memory around the requirements of threading episodic memory into every touchpoint. Our AI coach decides for itself when to reach back into what it knows about a person. To see where it actually does, we split sessions by whether it drew on prior context and compared the kinds of conversations on each side. Developmental sessions were 28 percentage points more likely than the average session to draw on prior context. Conversations where the coach acted as a thinking partner were 19 points more likely, and conversations about a specific situation were 16 points more likely. One-off operational tasks went the other way: 16 points less likely to draw on memory, and direct tell-me or do-this requests were 19 points less likely.
BetterUp AI coach sessions • Drawing on prior context
Where the coach reaches back
Difference from the average session, in percentage points, in how likely a session is to draw on prior context. Bars to the right draw on memory more often.
View the data as a table
| Session type | Difference (pts) |
|---|---|
| Developmental (growth over time) | +28 |
| Thinking partner (working an idea through together) | +19 |
| Live situation (a specific moment) | +16 |
| Operational, one-off (single-sitting tasks) | −16 |
| Tell me, do this (direct requests) | −19 |
This is exactly the behavior you'd want from a good human coach. Memory still plays a part in quick, practical requests, but a simpler one. When you need help tightening an email 10 minutes before a meeting, it helps that the AI coach already knows who the email is going to and which project it's about. When you're working through a stalled career conversation or a relationship with your manager that's deteriorating, memory does a very different kind of work. The AI coach draws on the arc of what you've been through and on patterns it’s noticed in how you tend to handle pressure or conflict. It can connect today's situation to earlier ones you may never have linked to yourself, all from the rich context it’s built up about you over time. This is what separates remembering someone from keeping a file on them.
Henry’s tape for Lucy was an act of love. It let her live a fuller life than she otherwise could’ve, but nobody would mistake it for her actually remembering. The gap between a record and an understanding is where the real work of AI coaching memory lives.
An AI coach needs more than a bigger store of facts. Chatbot memory records what you’ve said. Coaching memory has to understand what your experiences meant, how you’ve changed and when bringing something back will help rather than get in the way. That’s why our AI coach reaches back into prior context most often in developmental conversations and least often in quick, one-off tasks, which is the same judgment a good human coach uses. The goal isn’t to keep a file on someone. It’s to remember them.
Up next: Remembering on purpose: Inside our AI coach's memory
See how BetterUp’s AI coach remembers each person’s growth, not just their facts. Talk to BetterUp