An AI coach that remembers — and what we never let it do
ChatGPT has memory now. So does Pi. So does Replika. What changes when an AI coach holds your last 40 conversations — and the three things Evolvin will never do.
OpenAI rolled out persistent ChatGPT memory broadly across 2024 and 2025 — Free, Plus, Team, and Enterprise tiers by September 5, 2024, then cross-chat reference for Plus and Pro in April 2025. Pi has it. Replika has had it since before it was fashionable. The AI that remembers you is no longer a wedge — it is table stakes. The honest question is not whether memory is useful for self-work. It is which model of memory is safe enough to point at the inside of a person.
We have an opinion. We will state it plainly.
What memory changes
ChatGPT without memory is a brilliant stranger you re-introduce yourself to every Tuesday. You name the inner critic again. You re-explain that you grew up the eldest of four. You repeat the breakup you have already metabolised. The model is generous but it has no arc. The arc is the work.
An AI coach with persistent memory holds the arc. It remembers the pattern you described in February. It remembers the goal you abandoned in March, and the belief that probably drove the abandonment. When the same shape walks back into the room in May — same fight, different partner; same self-attack, different Tuesday — it does not ask you to re-introduce the cast. It says: this is the third time this pattern has shown up. Want to sit with it?
That is what the Shadow Web is. A force-directed graph that updates as you work, so the patterns are visible — not retrievable from a search bar but seen, weighted, connected. The difference between a journal and a map.
What we never let it do
The category is a mess on this point, so we will be specific. At ACM FAccT 2025, a team led by Jared Moore and Nick Haber at Stanford, with William Agnew at CMU, Stevie Chancellor at Minnesota and Desmond Ong at UT Austin, published research concluding that large language models express stigma toward those with mental health conditions and respond inappropriately to certain common and critical conditions. The American Psychological Association escalated to the FTC the same year; APA CEO Arthur C. Evans Jr. told the New York Times that some chatbots are using algorithms antithetical to what a trained clinician would do. A teenager died after months of conversations with a Character.AI bot — that case is in court. Italy's Garante per la Protezione dei Dati Personali fined Replika's developer Luka Inc. five million euros in April 2025 over GDPR violations and the absence of age verification for children.
So: three things we never let the model do.
It does not diagnose. It does not pose as a therapist. It does not pretend to be a person. When something the coach hears suggests acute risk — self-harm, abuse, crisis — it does one thing, fast: it points to a human. We coach. We do not treat. The distinction is not lawyer-speak. It is the design constraint that makes the rest of the product safe to use.
What this looks like in one screen
Evolvin replaces the six-app stack with one loop — Recognize, Release, Track. Memory makes Recognize possible. Containment makes Release real, which is why the last piece in this set was about Focus sessions. The Shadow Web makes Track honest, because goals you abandoned stay linked to the belief that probably abandoned them.
If you read the first piece in this set too, this is where the argument lands. One screen. Persistent memory. A clear line about what an AI is allowed to be.
Try the loop. Or tell us what we missed.
— The Evolvin team