Conversations & memories

Omi keeps two kinds of records. Conversations hold everything about one stretch of talk. Memories are short facts about you, extracted from conversations and linked back to them.

​

Processing a conversation

When a conversation ends, the backend:

1. Generates a title, an overview, and a category from the transcript.
2. Extracts action items and calendar events.
3. Extracts memories worth keeping about you.
4. Indexes the conversation for search and runs any apps you installed that act on finished conversations.

Conversations that turn out to have no useful content are marked discarded rather than deleted.

​

Conversations

Uploads that arrive out of order are merged by capture time, so a conversation recorded offline lands where it belongs. Very long recordings are split at segment boundaries into several conversations, each with its own summary.

​

Recurring voices

At batch finalization, confident unnamed voices can contribute to private profiles scoped to your account, under the existing voice-retention settings. Recurrence can associate later speech with the same unnamed voice; it does not create a person or assign a name. Confirming a person through an existing speaker assignment can authorize future automatic naming. The additive segment field recurring_voice_id carries that association without changing person_id before confirmation. The live speaker roster stays separate. Retention limits, deletion behavior and rollout controls are documented alongside the backend in ADR-0008.

​

Memories

A memory is one short fact, such as a preference, a habit, or something you mentioned. Each memory records the conversation it came from, so you can always trace it back. Extraction is deliberately selective: it skips mundane details and near-duplicates. You can review, edit, and delete memories in the app.

​

Reading this data

Conversations, memories, action items, and goals are available through the Developer API, the MCP server, and the CLI, all scoped to your account.