Knowing what to leave behind: Forgetting, consolidation, and visibility

This is the third 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 one, on why an AI coach needs a different kind of memory than a chatbot, and part two on how four types of memory were built into our AI coach.

For more than a century, forgetting has been treated as a failure: our memories fading away. But what if, as two neurobiologists argued in 2017, forgetting is actually part of how memory works well? Forgetting helps memory generalize, so we draw lessons from experience without overindexing on any single moment.

The idea of curation — what to include and what to exclude — lands differently in AI coaching. A system that holds so much information about a person has to be as thoughtful as a good human coach about what it brings back in conversations and what it lets fade away.

For example, let’s say someone vents to their coach about a colleague after a bad afternoon, and a month later the coach treats that colleague as an adversary. Or perhaps someone works toward a promotion for months, then decides that they don't want it anymore, and the coach keeps asking how the promotion plan is going. (If you’ve used an AI assistant that keeps resurfacing a project you dropped weeks ago, you know the feeling.) In both cases the coach remembered accurately, but used what it remembered badly.

Human coaches learn to avoid these missteps through training and supervision. A good coach holds what you told them in a vulnerable moment with care and waits for you to raise it again or for the right moment to revisit the subject. They also know that your account is only one side of the story. BetterUp’s research on more than 1,600 conversations between strangers found that people regularly misjudge how the other person experienced the same conversation, often underestimating how much their partner liked them.

We've tried to build the same instincts into our AI coach. A person’s account of a conflict is how they experienced it, not the full picture. Our AI coach is built to hold that nuance and help people see what’s underneath the experience and to challenge them rather than simply agree. This is a big part of what sets our AI coach apart from general-purpose chatbots.

Consolidation and retrieval: The work that happens between conversations and in time

Neuroscientists have shown that the brain replays and reorganizes the day's experiences during sleep, folding new episodes of what happened into what you already know. Our AI coach does its own version of this. After a conversation ends, or when new data arrives from an integration, agents work through it in the background, connecting it to what came before and noticing patterns in how the person tends to handle pressure or conflict. By the time the member comes back, or better yet, the next time our AI coach proactively nudges the user before a high-stakes meeting on their calendar, the AI coach's understanding is already current, and a pattern that took three conversations to emerge is ready to use in the next one.

Of course, all of that work would be wasted if the coach couldn't reach the information at the right moment. To do this, the AI coach sends out “search agents” to pull relevant context and ask questions of its memory the way a thoughtful person would, scoped in time and pointed at change. For example, how has the way this person talks about delegation shifted over the last two months? What did they commit to the last time their manager came up? Then it synthesizes an answer from what it finds.

Think back to the first-time manager in part one. Two months into her coaching journey, she opens a conversation with "I think I'm getting better at hard conversations, but I'm not sure." The AI coach can look across the conversation where she rehearsed telling her engineer his promotion case wasn't ready, the week he went quiet, the peer conflict she worked through in between, and the feedback in her latest 360. It can then reflect back what has actually changed in how she handles these moments, and what hasn't yet.

The bar for bringing this kind of history into a conversation is deliberately high. The AI coach does reach back into past conversations on its own, but it’s selective about when and how. Something shared in a hard moment informs how the coach supports the person, and it comes back into the conversation only when doing so will help. Memory does its deepest work in developmental conversations and a lighter, more practical job in quick tasks.

Memory in the brain isn’t fixed, either. Neuroscientists have found that recalling a memory can reopen it to change before it’s stored again. As new conversations and new data come in, agents update what the AI coach understands, weighing recent evidence more heavily than old patterns. For example, someone who needed a lot of reassurance before hearing hard feedback six months ago may want it straight now. Older episodes stay in the record as part of the person's story, marking how far they've come, but aren’t a description of who they are today.

Visibility: Memory you can see

An AI coach's memory shouldn't be a black box. When someone new comes up in a conversation, the user sees a card for that person showing what the coach now knows, and can review or edit it. To test how effective this visualization would be, we ran an experiment; people who were shown the card were 21 percentage points more likely to come back within seven days.

It would’ve been easy to assume that showing people what an AI remembers about their colleagues might make them more guarded. But the opposite happened. My guess is that trust grows when people can see what's being held on their behalf.

For the organizations deploying our AI coach, memory comes with safety guarantees. It only draws on what you feed it, so external sources, such as HR systems and Work IQ, only work when a company has authorized those integrations. Customer data is never used to train external frontier models, a restriction our model providers are held to by contract. Coaching conversations and memories are used to personalize each member’s coaching, and, only in de-identified and aggregated form, to improve our AI coach. And what an individual shares stays with them: those conversations and memories are never visible to their employer.

Lucy's tape in 50 First Dates could only ever tell her who she’d been. An AI coach has a harder job, because it has to hold on to someone's history while leaving room for who they're becoming. It earns the right to remember people by being careful with what they've trusted it with. That nuance, much more than the volume of what it can recall, is what we built our AI coach's memory around.

See how BetterUp’s AI coach remembers each person’s growth, not just their facts. Talk to BetterUp