How Echo remembers you without the cloud remembering you
EchoSense gives its AI companion long-term memory — interests, sensory notes, calming preferences — in encrypted on-device storage with semantic deduplication.
A companion that forgets you every morning isn't a companion. But the usual way AI products 'remember' is to store your life on their servers. EchoSense refuses that trade: Echo has long-term memory, and every byte of it lives encrypted on the user's own iPhone.
What Echo remembers
- Interests — the special topics that light a user up, so conversation starts warm instead of cold.
- Sensory notes — sounds, textures, and situations that overwhelm, so Echo never suggests the wrong environment.
- Calming preferences — which breathing pattern, which story persona, which music actually helped last time.
Two-pass semantic deduplication
Naive memory systems bloat: 'likes trains', 'loves talking about trains', 'is interested in railways' pile up as separate facts. Echo runs a two-pass semantic deduplication — new candidate memories are compared by meaning against existing ones, then merged or discarded. The memory stays small, current, and fast to inject into the model's context.
The storage stack
Memories live in Hive with encryption at rest; structured data lives in Drift (SQLite); keys live in the iOS Keychain. Deleting a memory in the UI deletes the data — there is no server-side copy to linger in a backup.
The cloud backend handles accounts and caregiver linking only. It has no table for conversations and no table for memories — a leak of our servers would expose no one's inner life. That is what privacy by architecture means: the sensitive data isn't protected by policy, it's absent by design.
Questions about EchoSense, pilots, or partnerships?
Talk to the team