Data retention
Data Retention sets how long LLM Gateway, MCP Gateway and OpenAI Compliance data is kept. Data past its retention period is deleted. You set a default horizon for each kind of data, then add ordered rules for traffic that needs a different horizon.
- Deletes data past its retention period. Once data passes its horizon, it is removed from the console and deleted.
- Applies to existing data. The horizon is measured from each event's time, so data collected before you published is covered too.
- Applies across the console: Overview, Costs & Savings, Graph, Users, Findings & Interactions, dashboards and Export Center exports created afterwards.
- Files already downloaded are not affected.
Summary tiles
Like the Policy Engine, retention uses a shared draft: edits are staged, and publishing creates the next revision.
Default horizon
How long each kind of data is kept when no rule matches. Each kind has its own duration, from 30 days up to 10 years.
Shortening the horizon for Request content and Response content while keeping Activity metadata longer is a common way to keep usage trends without keeping prompt text.
Ordered rules
Rules give specific traffic a different horizon. An event takes the first rule that matches; whatever is left falls to the default.
- Add rule to create one. Order matters, so put the most specific rules first.
- The default rule, Everything else, cannot be removed. It catches every event no rule above claimed. It has a Preset (for example Full activity) and a duration per data kind.
Review and publish
Edits are staged in a draft. The Review and publish section shows what is staged, for example "Publishing draft 4 replaces revision 3 everywhere in the console".
- Save draft to stage your edits. You must save before you can preview or publish.
- Validate checks the draft for errors.
- Preview impact shows, per data kind, how many events and characters the draft would affect and their share of the total.
- Publish makes the draft the live revision. Publishing applies the policy to everything already collected, not just to new data.
Verify a published policy
- Wait until the In force tile shows the new revision with Enforcing rather than Applying policy.
- Open Findings & Interactions and look at an LLM or MCP interaction older than the horizon you shortened. The data kinds you shortened, for example request content, should no longer be shown.
- Compare with Preview impact, which estimated how many events each data kind would affect.
Discard draft drops the staged changes and keeps the live revision.
History
Published revisions lists every revision with its number, publish time, rule count and who published it. The live revision is marked Enforcing. The first entry may show Baseline instead of a time. History records configuration, actor and time only, never the data itself. There is no one-click rollback: to go back, edit the draft to match an earlier revision and publish it.
What retention does and does not affect
Files you already downloaded are not affected. Handle them under your own retention process when they must follow the new policy.