---
title: Caching
description: How Agentside caches rendered twins in memory and how to tune or disable the cache.
sidebar:
  order: 6
---

Rendered twins are kept in an in-memory Effect cache so repeated reads skip your render function and the network.

```ts
agentside({
  routes,
  cache: { ttlMs: 60_000, capacity: 5000 },
});
```

| Option | Default | Description |
| --- | --- | --- |
| `cache.ttlMs` | `300000` | How long a successful render stays cached, in milliseconds. |
| `cache.capacity` | `1000` | The maximum number of cached twins. |
| `cache: false` | | Render on every request. |

## What is cached

- **Successes only.** Errors and missing twins expire immediately, so a failing source is retried on the next request.
- **Per path and variant.** `/pricing` for the control and `/pricing` for `pricing-openai` are separate entries. The request itself is not part of the key.
- **Concurrent requests share one render.** When ten agents ask for the same uncached twin at once, the source runs once and all ten get its result.

The `cached` field of each [`ServeEvent`](/docs/variants#exposure-events) tells you whether a response came from the cache.

## Output that depends on the request

Because the request is not part of the cache key, the first request to a path decides what the next ones get until the entry expires. When a source reads headers, cookies or the agent to build its output, model each output as a [variant](/docs/variants) or set `cache: false`.

## HTTP caching

The in-memory cache lives in one process. Markdown responses also carry `Cache-Control: public, max-age=3600` and `Vary: Accept, User-Agent` so a CDN can cache them safely. Change the header with `cacheControl`.

```ts
agentside({
  routes,
  cacheControl: "public, max-age=300, stale-while-revalidate=3600",
});
```
