Context cache
With the context cache, the model provider briefly keeps the part of a conversation it has already processed. Replies start faster and use fewer credits. Admins decide whether Custodos requests a cache and for how long.
Last checked on 28 September 2026
On this page
What the cache does
With every new message, the model receives the whole conversation so far again, including instructions and attached files. Without a cache, the provider processes this text from scratch every time. With a cache, it keeps the part it has already processed for a while, and the next message builds on it.
The cache sits with the same provider and in the same region that processes the prompt anyway. The processing region does not change. No additional content is stored, only intermediate results for text the provider receives anyway.
The settings
| Setting | What applies |
|---|---|
| Off | Custodos requests no cache. Every message processes the whole conversation again. This costs more credits, and replies start more slowly. |
| Standard | Five minutes, renewed with every message. Suits a conversation held in one go. Default for new workspaces. |
| Extended | A longer duration, so a conversation survives a break. Only takes effect on models for which Custodos has confirmed the longer duration with the provider. In the current model catalogue no model is confirmed yet; until then, the Standard duration applies. |
Changing the cache duration
- 1
Open the settings
As an admin, click Settings and go to the Context cache section. - 2
Choose the duration
Under Cache duration, choose Off, Standard or Extended. The choice is saved immediately and recorded in the activity log.
Kept separate per workspace
Custodos tags every request so that a cache entry can only be used by requests from the same workspace. Entries from another workspace cannot be reached by your requests, and the other way round.
When the data protection setting, a chat's security mode or access to a brain changes, a new cache starts. Older entries can then no longer be reached and expire at the provider on their own. They cannot be deleted earlier, because no provider offers that.
When the setting does not apply
- Some models cache on their own and cannot be told not to, in their own region and on the same data they process anyway. For these models, the provider's own duration applies. Off therefore means that Custodos requests no cache.
- In security mode, Standard is the maximum, even if the workspace has chosen Extended.
- Extended only takes effect on models for which Custodos has confirmed the longer duration. As long as none is confirmed, Extended behaves like Standard.
- Very short conversations fall below the provider's minimum length. No cache entry is created then.
Read next
Still have a question? Write to us.