We built an AI that forgets on purpose
The entire industry is pulling in one direction: remember more. Bigger context windows. Permanent vector stores. "Memory" as a checkbox on the pricing page. The pitch is always the same — your agent will never forget anything you tell it.
We built ARIA the other way. Our agents have a nightly prune job that runs on a Temporal cron and deletes memory on a schedule. Working memory expires in hours. We spent real engineering time teaching the system to let go.
That sounds like a step backward. It is the most important design decision we made.
Perfect recall is a liability
A sales agent that perfectly remembers every fact it has ever seen is not smart. It is cluttered. Last quarter's pricing. A prospect who churned. A competitor feature that shipped and then got pulled. A "preference" the user stated once, sarcastically.
Human memory is not a hard drive — it is a ranking function that aggressively throws away what stopped being true. The reason you can act decisively this morning is that you are not carrying every contradictory thing you learned over the last two years. Agents that hoard get slower and more confidently wrong, because retrieval starts surfacing stale facts next to fresh ones with no signal about which is which.
The failure mode is specific and brutal: an agent confidently acts on a fact that used to be correct. Infinite memory doesn't prevent that. It guarantees it.
What forgetting actually looks like in ARIA
We don't have one memory bucket. We have four, and they decay at different rates:
- Working memory — what an agent is holding for tonight's run. Gone by morning.
- Episodic — "this specific thing happened on this date." Useful, but it ages out; an event from six months ago shouldn't weigh as much as last night's.
- Semantic — durable facts about the business. These persist, but they carry a confidence score (we default semantic facts to 0.7, not 1.0) so nothing is treated as gospel.
- Procedural — learned ways of doing a task. The most valuable, the slowest to change.
Every memory carries metadata, an audit trail, and a soft-delete flag. Nothing is hard-deleted on the first pass — a memory is marked, decayed, then pruned. We can always see why something left.
The result: when an agent retrieves context before acting, it is pulling from a set that has already been weeded. Recent and durable rises. Stale and one-off sinks and eventually leaves.
Forgetting is a confidence problem in disguise
Here's the part nobody talks about. The hard question isn't "what should we delete." It's "what changed."
When a new fact contradicts an old one, you have a conflict, not a duplicate. Naive systems either keep both (and confuse retrieval forever) or blindly overwrite (and lose the history). We do neither — we detect the conflict, keep the audit trail, and let the newer, higher-confidence fact win the retrieval, while the old one decays out rather than vanishing instantly.
That's why confidence and forgetting are the same system. A fact you're 0.7 sure about and haven't seen confirmed in months should quietly lose to a fact you saw last night. Decay is just confidence with a clock attached.
The uncomfortable takeaway
If your agent's memory only ever grows, you are not building memory. You are building a landfill with a search bar.
The agents that will actually be trusted to act overnight — to email a customer, to adjust a spend cap, to file a task as done — are the ones that know what is currently true, not the ones that remember the most. That requires the unglamorous machinery: TTLs, decay curves, conflict detection, a prune job, an audit trail for every deletion.
We built an AI that forgets on purpose because the alternative isn't a smarter agent. It's a more confident one. And in autonomous work, confident-and-stale is the most expensive failure there is.