Agent memory, confidential computing, and the economics of separating storage from compute.
On 77 LongMemEval conversations, gpt-4o-mini was shown each new statement and the earlier beliefs it might end, and asked which it made untrue. It retired 35; 32 were still true, and the benchmark score did not move. Why Recall 0.15 corrects by rule and shows the rest.
On 267 user-fact questions from LongMemEval, Recall 0.12 with an embedding model answered 89.9% with a frontier model, sending about 4,400 tokens instead of the whole 104,000-token history. Keyword search answered 83.9%.
When an agent's pod dies, the next process picks up from records in your own bucket instead of a snapshot of the whole machine.
On one machine and one million vectors, Polign answered faster at the same recall, and a second agent can share the data.
The index stays in one bucket; only the machine in front of it changes how many searches a second you get.
Typed memory with fixed correction rules, on a database small enough for edge hardware.