Retention, TTLs, and legal holds
Shared by Pavel · Updated 2026-11-12
A sound policy should set explicit expiry by data class, distinguish ordinary memory from retained safety or legal records, and make any deletion pause narrow, auditable, and time-bounded.
Expiry and retention
More: TTL policy and retention periods
- Use a time-to-live (TTL) to automatically delete memories after a defined period; one documented approach is to expire memories that have not been accessed for a long time.
- Set retention separately for user-facing memories and records kept for safety, review, or legal reasons: deleting a memory for the user need not mean those records are immediately deleted.
- Make deletion cover derived copies as well as source data; use tombstones to propagate deletion to replicas or derived copies.
Holds and the OpenAI example
- A legal hold overrides ordinary retention rules and a user's delete request for a defined set of records; it should be logged and should end when the hold ends.
- In May 2025, a court ordered OpenAI to preserve output-log data that otherwise would have been deleted.
- On 9 October 2025, the court lifted the ongoing preservation duty. Logs already saved and data on accounts flagged by The New York Times remained preserved.
- OpenAI later said its preservation obligations ended on 26 September 2025. The orders show that a preservation pause can stop routine deletion, but lifting the forward-looking duty does not itself erase records already preserved; the scope and end of the hold matter.