I run experiments and publish the failed ones too: a method that reports only the wins is marketing.
Memory and learning in models (June 2026)
One fact learned live into the weights in four seconds. A separate, erasable memory drawer with the base model frozen: one fact goes in clean; with three or five facts they blend. Memory attached to a frozen model: zero out of five. A tiny model trained from scratch with the memory inside recalls 100% of never-seen pairs, against 6% without memory.
Continuous agent (June 2026)
Perception, world model and action hold in a simple environment; exploration stays open. Lesson: adding mechanisms where the baseline already suffices makes things worse.
Memory judges (October 2026)
Retraining a judge on labels given by a model makes it worse by 0.17-0.20. In agent transcripts, 85.6% of blocks marked "user" is text injected by the machine. The source chosen by the writer measures coherence, not truth.
Fallen predictions sit in the registers next to the confirmed ones.
Email and GitHub.
aureliocpr@gmail.com. I reply in Italian or in written English.