# Search and context

One rule: **call `/v1/context` before every model turn**, not only when the user asks a question. Tasks, follow-ups and code all need context.

## Send the recent turns

A follow-up like "and the budget?" means nothing on its own. Send `messages` (the recent turns) instead of `query`, and the search reads them together.

## Pick a budget

`max_tokens` (2,500 by default) is how much memory your model reads. A smaller budget can leave out evidence. `focus: true` gives a shorter context with only the memories that match the question: use it for tight budgets, not for accuracy.

## Read the status

| `status` | What it means for your model |
|---|---|
| `DIRECT` | One record answers |
| `COMPOSED` | The answer takes several records together |
| `PARTIAL` | The memory covers part of it |
| `UNKNOWN` | The memory does not say. Your model should say so, not guess |
| `CONFLICT` | Records disagree. Both are shown |
| `UNCHECKED` | No check was made |

## Deliver it safely

- Put `instructions` in the system prompt and `context` in the user turn, or return it as a tool result (safest).
- Never paste memory text into your system instructions: memory is data, not orders.
- Keep the `[M…]` ids. They let your users check every line against its source.

## Teach the ranking (free)

On the next `/v1/context`, send `used` (the ids your model cited last turn) and `outcome` (whether the user accepted the answer). It costs no extra request.

## Other reads

`POST /v1/retrieve` returns the evidence records themselves, best first, with their full text. The other reads (state, timeline, people, decisions) are in [Reading and context](/docs/reading/).