Memory belongs to the person, not the model
Why we think AI memory should be local-first and work across tools, and what that changes about how it's built.
Every AI tool is adding memory. ChatGPT remembers facts about you. Claude has Projects. Cursor keeps rules. Each is useful, and each lives inside one product. Switch tools and you start over.
We think that's the wrong shape. The context that matters — what you're building, what you've decided, what you ruled out and why — belongs to you. It should follow you between tools, and it should sit somewhere you control.
What memory is for
Most of what you say to an AI doesn't need to be remembered. A good memory system is mostly about forgetting: keeping the decisions, constraints and numbers, and letting the rest go.
The hard part isn't storage. It's retrieval — bringing back the three moments that matter for the question you're asking now, not the whole transcript. Too little context and the model guesses. Too much and it drowns, or anchors on something stale. The useful unit is small: one decision with its date, one constraint with its reason.
Why local-first
Personal context is the most sensitive data most people will ever give software. It includes half-formed plans, money, health and other people. We don't think it should default to someone else's server.
So our starting assumptions are strict. Speech is transcribed on-device. Memory lives in a local database. Sync is opt-in. Capture stays off until you allow specific apps, and password managers, banking, health and messaging are blocked outright. These constraints make the engineering harder. We think they're the price of being trusted with memory at all.
What we're testing
Murmur is where we test these ideas against real work. The open questions are practical: how often retrieved context actually helps, when it gets in the way, how people want to correct what the system remembers, and what should expire. We'll write up what we learn here.