Research note · Design
Local-First Memory for Agents
What an agent remembers can be more sensitive than the transaction itself. This note examines memory that stays on the user side.
A useful agent remembers: positions, risk appetite, recurring intents, corrections you have made. That longitudinal memory is more revealing than any single prompt, yet agent products routinely centralize it. Local-first memory inverts the default — memory is encrypted on the user's device, keyed to the user, exportable and deletable, with optional end-to-end encrypted sync for multi-device use.
The hard parts are the unglamorous ones: key recovery when a device is lost, migration between runtimes, memory poisoning via injected content, and proving to a skeptical user that nothing left the device. A model still has to read a prompt to answer it — local memory changes where the archive lives, not the physics of inference.
This note is marked Design rather than Research because a browser-local prototype is within reach of a small team, and building it would convert several claims here into demonstrable behavior.
What this note covers
- Encrypted local storage with user-held keys
- Export, delete and selective recall
- Optional end-to-end encrypted sync
- Failure modes: device loss, key loss, memory poisoning