---
title: Chats and History
description: How a bot chat lasts indefinitely, what the bot sees on each turn, and how it finds earlier messages.
sidebar:
  order: 2
---

Each bot has one bot chat with you, and it never starts over. There's no "new
chat" button: you keep talking in the same place, whether the last message was
a minute ago or a month ago.

Messages from [Discord](/concepts/discord) join the same bot chat. Only the
answer associated with each Discord request returns to Discord.

## Transcript and Working Context

A chat that lasts for months can't fit in a model's context, so Crumble keeps
two things apart:

| Name                | What it holds                                                                                                             | Size           |
| ------------------- | ------------------------------------------------------------------------------------------------------------------------- | -------------- |
| **Transcript**      | Every message, reply, tool call, and piece of background work.                                                            | Grows forever. |
| **Working context** | What the bot reads when it answers: recent messages, summaries of older ones, its instructions, and anything it recalled. | Kept small.    |

When the working context gets large, the bot **compacts** it. Older messages
are replaced by a summary in the working context, and the transcript stays
exactly as it was.

```mermaid
flowchart LR
  subgraph Transcript["Transcript (permanent)"]
    direction LR
    M1["Message 1"] --> M2["..."] --> M3["Message 400"] --> M4["Message 401"] --> M5["Message 402"]
  end
  subgraph Context["Working context (what the bot reads)"]
    direction TB
    I["Instructions"]
    S["Summary of messages 1-400"]
    R["Messages 401 and 402"]
  end
  M3 -. "compacted into" .-> S
  M4 --> R
  M5 --> R
```

**Compaction never deletes anything you can see.** It only changes what the
bot reads on its next turn.

## Images

Attach images to a message, and a bot whose model reads images sees them as
they are. A bot whose model can't read images refuses a message with an image,
unless the account has a [vision model](/getting-started/providers#vision-model).

With a vision model set, the vision model reads each image and describes it,
including any visible text. The bot works from the description, and the
transcript keeps the original image. The bot can also ask the vision model a
question about an image it already received, such as "What's the total on this
receipt?", so the first description needn't hold every detail.

If the vision model can't read an image, the bot gets a note saying so in its
place and answers anyway. The next turn tries the image again.

## Bringing Back an Earlier Message

Reply to a message to bring it back into focus. The bot reads that message and
the messages around it before it answers.

## Searching the Transcript

Bots can search and read their own transcripts, including messages that were
compacted away long ago. Ask in plain language, such as "What did I say about
the domain renewal in March?"

A bot searches only what it's allowed to see:

| Where you ask | What the bot searches                              |
| ------------- | -------------------------------------------------- |
| Its bot chat  | That chat, and the group chats the bot belongs to. |
| A group chat  | Only that group chat.                              |

In a group chat, a bot can't search its bot chat with you, even if you ask it
to. For how group chats work, see
[Group Chats](/concepts/group-chats).