# The context window

> Every model has a fixed amount of room. The indicator shows how much of it this conversation is using, and the panel shows what is taking the space.

A model reads a fixed amount of text per request — its **context window**. Everything the agent needs has to fit: its instructions, its memory, the tools it can call, your conversation, and your attachments.

The composer shows how full that window is, and the context panel shows what is filling it.

## Why it matters

Two things degrade as the window fills:

- **Older messages get trimmed.** The agent stops seeing the start of a long thread.
- **Every step costs more.** The whole context is re-sent on each step of a [run](/docs/work/runs/), so a heavy conversation is a slower and more expensive one.

An agent that "forgets what we said at the beginning" is almost always a full window, not a broken memory.

## What is taking the space

The panel breaks the window into segments:

| Segment | What it is |
| --- | --- |
| **System instructions** | The frame Endue puts around every run |
| **Identity** | The agent's name, character, and profile |
| **Agent instructions** | Its [system prompt](/docs/build/system-prompt/), at the current revision |
| **Instruction skills** | Skills whose content is standing guidance rather than a command |
| **Memory** | The [memories](/docs/capabilities/memory/) relevant to this conversation |
| **Tool definitions** | The schema of every tool the agent can call |
| **Conversation messages** | What you and the agent have said |
| **Attached media** | [Attachments](/docs/work/attachments/), counted separately from messages |

Tools are broken down further — each skill, connector, built-in group, and workspace shows what its definitions cost to advertise, and how many times it was called in the last run with what the results weighed. That is the difference between *"this tool is expensive to offer"* and *"this tool returned a lot"*.

## Measured versus projected

Before the first [run](/docs/work/runs/) there is nothing to measure, so the figure is a **projection** of what the next request would carry. After a run it is what that request actually contained.

The panel says which one you are looking at. A projection that differs from the measurement afterwards is normal — the agent may pull in a memory or a tool result that was not predictable.

## Resources

The panel also lists the attachments loaded in this conversation. Selecting one scrolls to the message it came with, which is the quickest way to find *"the screenshot from earlier"* in a long thread.

## When the window is full

In rough order of effect:

1. **Start a new [conversation](/docs/work/conversations/).** The single most effective move — a fresh thread carries none of the old messages.
2. **Move durable facts into [memory](/docs/capabilities/memory/) or the [project](/docs/capabilities/projects/)** before you do, so the next thread does not need re-explaining.
3. **Unbind tools the agent does not use.** Their definitions are re-sent on every step whether or not they are called.
4. **Pick a model with a larger window** — the picker shows each model's size. This buys room; it does not make a wandering thread focused.

<Aside type="tip" title="Long threads are not free even when they fit">
  The cost of a step is the whole context, not just your new sentence. Two focused
  conversations usually cost less and answer better than one that has covered
  three subjects.
</Aside>

## Limits

- The window size comes from the model. Changing model changes the ceiling.
- Trimming drops the oldest messages first; it does not summarize them.
- The breakdown reflects the last run. Changes you make afterwards — binding a tool, editing the prompt — show up on the next one.

## Related

<CardGrid>
  <LinkCard
    title="Conversations"
    href="/docs/work/conversations/"
    description="When to start a new thread instead of continuing."
  />
  <LinkCard
    title="Memory"
    href="/docs/capabilities/memory/"
    description="Carrying facts across conversations instead of in them."
  />
  <LinkCard
    title="Choosing a model"
    href="/docs/build/models/"
    description="Window sizes, and what each model costs per token."
  />
</CardGrid>