Core concepts
Long-term memory
What the agent remembers across conversations, who decides, and how memory travels when work leaves the device.
Short-term memory is the conversation. Long-term memory is short notes the agent keeps across conversations, in the device's encrypted store. Three rules: nothing is remembered without your approval, forget means forget, and scope decides where a memory may go — never whether it can be used on the device.
- 1
Notice
Each turn gets a near-free salience check ("remember…", "from now on…", corrections, schedules). A memorable moment becomes a proposal — the app shows it (
memory.proposed). - 2
Approve
Only approved notes are ever used. Proposals never reach a prompt.
- 3
Recall
Each turn, the most relevant approved notes are added to the prompt inside a small-model budget. Corrections and procedures become rules the agent follows; preferences, facts and events go in the memory section (
memory.recalled, traced). - 4
Forget
forgetMemorydeletes the note and leaves a tombstone so the agent doesn't quietly learn it again.
| Scope | Leaves the device? |
|---|---|
user_private (every note starts here) | never |
user_synced (shareMemoryWithMyNodes) | only to your own paired nodes, only for the kinds each node is granted |
org_shareable | to the organisation brain — redacted, and believed only when several devices agree |
const { pending } = await lybo.listMemories();
await lybo.approveMemory(pending[0].note_id);
await lybo.shareMemoryWithMyNodes(pending[0].note_id); // optional
const preview = await lybo.recallMemory("what should I pack for Lily?");
await lybo.forgetMemory(pending[0].note_id);When the LYBO Harness hands a task to another node, it builds a handover capsule: only shareable notes, only granted kinds, personal details removed, expiring, digest-signed. The node borrows it for one session and never stores it; what the node notices comes back to your device as proposals.