---
title: Personal context and memory
description: Follow recorded speech into Recordings, Notes, Memory and Feed, with distinct source and deletion rules.
---

Impo uses conversations, recorded speech and connected services as context. Each product surface uses that context differently.

## Recordings, Notes, Memory and Feed

| Surface | What it contains |
| --- | --- |
| Recordings | Original phone recordings and their transcripts, with anonymous speaker labels and speaker review. |
| Notes | Finished, read-only documents derived from the full speaker conversation, with links to their sources. |
| Memory | Long-term personal context consolidated from chat and the user's confirmed speech. |
| Feed | Useful findings from background checks, supported by personal context and available connected services. |

**Rewind** names the recording feature. **Recordings** names the original-capture view. Notes can use other speakers' words to describe a conversation; that does not make those words the user's personal statements in Memory or Feed.

## Rewind becomes personal context

The phone captures audio under native permission and lifecycle rules. Immutable batches upload with server-issued S3 URLs. Workers transcribe the batch into anonymous speaker turns.

The user confirms which voice is theirs. That selection produces a shared personal-transcript projection used by Memory and Feed. Unconfirmed speech is not silently treated as the user's own speech. Speaker labels do not identify the same person across recordings.

Changing the review or deleting the recording revises that projection and withdraws stale derived content. These are data dependencies, not just UI filters.

## Notes summarize one session at a time

Every Rewind recording is a session with a length chosen at start; it ends on its own at that length or when the user stops it. When a session has ended and all of its recordings are transcribed, the Worker makes one Agent call with the session's whole transcript and stores the one Note it returns, under the session's id. The Agent has no tools and no memory of other sessions: it orders the document by topic, cites the exact utterances, and never guesses names or dates.

A late upload, a speaker correction or a deletion marks the session changed. Its Note is hidden until the next run rewrites it in place. Notes lists only sessions whose Note reflects their current recordings; sessions still recording, still transcribing, or without speech show no Note.

## Memory and retrieval

Chat and confirmed Rewind context can contribute to long-term memory. Memory retrieval is part of the main agent's instructions. The store keeps source references and change history so forgetting and expiration can be applied consistently.

Adding a new context source means defining ownership, retention, source revisions and deletion behavior. Do not write arbitrary derived text directly into a prompt and bypass the store's lifecycle.

## Feed and connected apps

The hourly background step uses an isolated agent through the Agent API, with read-only access to connected cloud apps. It looks for useful, supported findings. Empty checks produce no visible edition.

Cards retain source resources and connector generations. Disconnecting a service revokes dependent cards, actions and pending notifications. Native phone tools are excluded from these unattended checks.

## Storage in the current implementation

Execution records and personal content have different lifecycles. A tool receipt records an operation; it is not long-term memory.

| Store | What it owns |
| --- | --- |
| PostgreSQL | Users, ownership, submissions, tool receipts, schedules, connector bindings, Notes and recording metadata. |
| Agent API provider (currently Rebyte) | Managed Sessions, Turns, conversation Items, sandbox working files and delivered artifacts. |
| S3 | Private audio uploads and per-user Rewind transcript records in the production configuration. |
| Turso | A separate long-term memory database per user, including content, history and embeddings. |
| Client storage | Account-scoped pending commands, recording batches, drafts and device receipts, according to platform. |

Local development can retain content in local PostgreSQL. It does not simulate production storage durability simply by exposing the same API.

## Extending the data model

Drizzle entities are the schema source of truth. Put new queries in a domain repository, keep complete transactions inside it, and use `npm run db:push` during development. Do not create a parallel migration history.

See the [repository boundaries](https://github.com/impoai/impo/blob/main/server/README.md#database-and-code-layout), [Rewind speaker contract](https://github.com/impoai/impo/blob/main/contracts/echo-speakers.md) and [Brief contract](https://github.com/impoai/impo/blob/main/contracts/brief.md) before adding a source or a derived product surface.
