> ## Documentation Index
> Fetch the complete documentation index at: https://docs.augent.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Memory & Caching

> How Augent stores transcriptions, embeddings, and diarization results so every search after the first is instant.

Augent remembers everything it transcribes. The first run processes the audio. Every operation after that is instant.

***

## Cache key

Every transcription is keyed by **file hash + model size**:

```
SHA256(file_content):model_size
```

* Same file, same model = instant cache hit
* Same file, different model = new transcription
* Modified file = different hash = new transcription

The file content is hashed in 8KB chunks using SHA256. This means renaming or moving a file doesn't invalidate the cache — only changing the content does.

***

## Storage location

Everything lives in `~/.augent/memory/`:

| File                  | Purpose                                                                   |
| --------------------- | ------------------------------------------------------------------------- |
| `transcriptions.db`   | SQLite database with all cached data                                      |
| `transcriptions/*.md` | One markdown file per transcription (human-readable, Obsidian-compatible) |

***

## What gets cached

| Data             | Cache key                   | Storage                             |
| ---------------- | --------------------------- | ----------------------------------- |
| Transcriptions   | `file_hash:model_size`      | SQLite + `.md` file                 |
| Embeddings       | `file_hash:embedding_model` | SQLite (numpy BLOB)                 |
| Diarization      | `file_hash:num_speakers`    | SQLite                              |
| Source URLs      | `file_hash`                 | SQLite                              |
| Audio separation | `file_hash:model:stem_mode` | Filesystem (`~/.augent/separated/`) |

Each type of data is cached independently. You can diarize with different speaker counts without re-transcribing. You can run semantic search without re-computing embeddings on subsequent queries.

***

## Model caching

Whisper models stay loaded in memory between tool calls. The MCP server is a long-lived process — once a model is loaded for the first transcription, subsequent transcriptions with the same model size are faster because there's no model loading overhead.

The sentence-transformer model (`all-MiniLM-L6-v2`, \~80MB) is also loaded once and kept in memory.

***

## Translations

When you translate a non-English transcription, the English version is stored as a sibling `(eng)` markdown file alongside the original. Both appear in the Memory Explorer and Web UI.

***

## Managing memory

| Tool            | What it does                                                   |
| --------------- | -------------------------------------------------------------- |
| `list_memories` | Browse all stored transcriptions with titles, durations, dates |
| `memory_stats`  | Total count, duration, and storage size                        |
| `clear_memory`  | Delete all cached data                                         |

Or use the [Web UI](/web-ui/overview) to browse, search, and delete individual transcriptions.