Most conversational AI systems have a short memory by design. A typical large language model retains context only within a single session — once a conversation ends, whatever was discussed is gone unless a user explicitly reintroduces it. This is usually framed as a privacy feature. We think it's more accurate to call it a limitation, and one that's holding the entire category back from feeling genuinely conversational.
Human relationships aren't built on isolated exchanges. They're built on accumulation — the small, unremarkable details that get carried forward from one conversation to the next without anyone having to consciously re-explain them. When a friend remembers that you mentioned a stressful week, or a difficult decision, without you bringing it up again, that recall is part of what makes the relationship feel real.
Most AI systems can't do this. Ask a typical chatbot about something you mentioned three conversations ago, and it has no idea what you're talking about. That's not a safety feature — it's just an architectural gap. We built Remy to close it.
Remy retains context across sessions using a layered memory architecture: short-term working memory for the active conversation, and a longer-term store that persists meaningful details across sessions. We deliberately avoided the industry-standard approach of treating memory as a simple retrieval index — a flat database a model can query. Instead, Remy's long-term memory is weighted by emotional salience: the system prioritizes retaining details that carried weight in the moment they were shared, rather than treating every sentence as equally worth remembering.
In practice, this means Remy is more likely to remember the name of someone you were worried about than the name of a restaurant you mentioned in passing. We think this mirrors how human memory actually works — imperfectly, selectively, and often more attuned to what mattered emotionally than to raw information density.
This approach comes with a real cost: it's harder to predict exactly what Remy will and won't remember, and harder to give users a simple, complete accounting of what's stored. A flat retrieval index is easy to audit. A salience-weighted memory system is closer to how people actually retain things — which also means it's less legible from the outside, including to us.
We think that tradeoff is worth it. A memory system a user can fully audit line-by-line doesn't feel like being remembered. It feels like being logged. We built Remy to do the former, even where it makes the latter harder to guarantee.
We're continuing to refine how memory decay works over long timescales — how much detail Remy retains a year into a relationship with a user, versus a week in, and how to make sure older details resurface at the right moments rather than at random. This is ongoing work, and we don't expect to have it fully solved for some time.