Memory is one of the things that makes Thoughtful feel different from talking to a general AI. When a user starts a session, the assistant remembers what they've talked about. They don't have to re-explain the situation at work, the friend they're worried about, the pattern they noticed last week. The context is already there.
It's a feature that sounds simple. Underneath, it's one of the more careful design decisions in the product. Memory is also a clinical responsibility, and a privacy commitment, and a craft problem. What you choose to remember about someone shapes the experience they have with the product. What you choose to forget matters just as much.
What we keep
After every session that meets a meaningful engagement threshold — four or more user turns — the system runs a structured set of jobs in the background. They generate a session summary with four sections: what the user came in with, how the conversation reframed the challenge, the concrete plan that came out of the session, and the patterns or insights the assistant surfaced.
They also produce two short pieces of metadata used in the home screen. A topic title, which becomes the headline on the Recent Topic card the user sees on the home feed. A short framing line, which previews what the next session could pick up on.
And they update a longitudinal log: which therapeutic techniques came up, which tools the user engaged with, and a timeline of the user's themes over time. This is what makes the product feel like it knows you.
What we don't
Memory isn't the full transcript played back. It's a summary that an assistant uses to orient. The reason this distinction matters is that what gets remembered shapes what gets reinforced.
If we kept everything — every digression, every false start, every momentary frustration — the assistant would be drawing on a noisier picture of who the user is than is actually fair to them. People say things in conversation they don't fully mean. They explore framings they later abandon. A mental health AI that takes every word as a fixed data point about the user is mis-modelling them.
So memory is summary-shaped. What was the user working on? What did they decide? What's the plan? What's the pattern? These are durable things. The exact phrasing of a specific message isn't.
The four-turn threshold
Some sessions don't generate memory at all. Specifically, sessions where a user has fewer than four turns of conversation.
This is a small line in our spec, and it does meaningful work. The reasoning: a session of two or three turns isn't enough material to draw a reliable summary from. It's enough to register that the user opened the app, started something, and stepped away. It's not enough to know what they were actually working on, or what to remember about it.
So short sessions don't get summarised. They don't generate Recent Topic cards. They don't appear in session history. The safety check still runs — that part is universal — but the memory layer respects the threshold.
This protects the user from a worse version of the product: one where every fragmentary interaction generates a slightly wrong impression of them that the assistant carries forward. A product with bad memory is worse than a product with no memory. The threshold is how we avoid building the bad version.
Two summaries, two jobs
Worth noting because it's a quiet piece of craft.
Two pieces of writing get generated at the end of a session. The Session Summary — the full four-section recap of what happened — is for the user. They can review it after the session. They can come back to it from history at any point. It's the artefact of what they did.
The Recent Topic card — the short title and framing line that appears on the home screen — is for re-engagement. It's the bridge from a finished session to the next one. It uses different copy from the Session Summary because it serves a different purpose: not 'here's what we talked about' but 'here's what we could pick up next.'
These look like two views of the same thing. They're not. They're two pieces of writing, generated by different background jobs, optimised for different moments. Conflating them would have produced a worse product.
How users actually use it
The most engaged users tell us something specific about memory. Habitual users start with a narrow concern in their first few sessions, and then progressively work across their whole life — relationships, work, sleep, money, self-image — as the assistant gets enough context to hold all of it at once.
Roughly 70% of sessions involve users navigating multiple life domains simultaneously. Memory is what makes that possible. Without it, every session would have to begin with a context dump. With it, the user can show up to the next session having already moved forward, and pick up where they left off.
This is also why the most engaged users describe the product as a place where they don't have to be selective. They can be messy. They can bring whatever is actually on their mind, even if it's only loosely related to the thing they thought they were working on. The product can hold the whole picture.
The principle
Memory is a feature, but the design principle is older than the feature. Remember enough to help. Forget enough to not weigh the user down. Keep the durable things — patterns, plans, themes. Don't keep the noise.
Most mental health products either have no memory, which makes every session feel like starting from scratch, or have memory in a way that records everything indiscriminately, which makes the product feel like it has a file on you. Neither of those is what we wanted.
What we built is closer to how a good clinician holds a patient over time. Enough memory to recognise the work in progress. Enough discretion to not reduce the person to the worst sentence they ever said in a session. That's the bar.