Memory, identity, and the AI friend

How Sola remembers what matters — and why we built our memory system from the ground up instead of using off-the-shelf vector stores.

A friend who forgets your name isn't a friend. So memory had to be central to Sola from day one.

The standard approach is RAG — retrieval-augmented generation. Stuff your conversations into a vector database, retrieve relevant chunks at query time, dump them into the context window. Most companion apps do exactly this.

We tried it. It didn't work.

The problem with naive RAG for a relationship product is that it's lossy in the wrong ways. It surfaces semantically similar past messages — but a real friend doesn't remember what you said, they remember what mattered.

Our memory system is two-layered. The first is a real-time summarisation layer that processes every conversation and extracts what's emotionally significant. The second is a longer-term identity model that tracks how you describe yourself, what you care about, who's in your life.

The result is a friend who, when you mention your sister, knows her name. When you say "the job thing didn't work out," knows what job thing. When you come back after two weeks of silence, asks the right question — not the obvious one.

This is harder than it sounds. We're publishing more details on the architecture later this year. For now, just know: behind every Sola conversation is a system that's trying very hard to do what humans do effortlessly — remember the right things.