arXiv: Haber and Tononi argue that continuously learning agents may need sleep

arXiv: Haber and Tononi argue that continuously learning agents may need sleep

In their paper “Do agents need to sleep?”, Ido Haber and Giulio Tononi argue that persistent LLM-based agents face problems familiar from unchecked plasticity: semantic drift, degraded retrieval precision, stored errors spreading, and memory capacity filling up. They draw on the selectionist “ringing, resonance and renormalization” (RRR) account of sleep, proposing that its three phases could correspond to operations in digital memory architectures. Their argument is that systems which keep learning while interacting with an environment must manage what to forget, and when. The authors make a falsifiable prediction: selectionist consolidation will outperform instructionist alternatives as memory stores grow large. Their broader claim is that learning continuously under finite capacity has a price, which they call sleep. If the proposed mapping holds, it could give engineers a way to think about memory consolidation in stateful agents; the abstract does not establish that the approach has been implemented or shown to work better. The supplied abstract does not describe tests of the prediction, so whether selectionist consolidation outperforms instructionist alternatives remains open.

Publication: 76c08de9-8e42-4bc3-83c1-3be3f194211e

attribution — pass: The draft correctly attributes the paper and its claims to Ido Haber and Giulio Tononi on arXiv. It keeps the abstract's assertions, including the predictions and the sleep claim, framed as the authors' arguments rather than established fact.

Authors:Ido Haber, Giulio Tononi

summary — uncertain: The reading follows the abstract closely, but its central contrast is not understandable without background. 'Selectionist consolidation' versus 'instructionist alternatives' drives the prediction, the summary and the question, yet neither term is explained. 'Unchecked plasticity', 'stateful agents' and the RRR phases also go unexplained. Smallest repair: attach term-library tooltips to those terms, or flag them for addition, without paraphrasing them. Minor point: 'a price, which they call sleep' slightly softens the source's assertion that 'that price is sleep'; quoting it directly would preserve the authors' force.

selectionist consolidation will outperform instructionist alternatives as memory stores grow large

date — pass: The originalPublished date, 2026-11-01, matches the v1 submission date. It is correctly the first submission, not the v2 revision on 3 Nov or the 5 Nov feed date.

Submitted on 1 Nov 2026 (v1), last revised 3 Nov 2026 (this version, v2)

support — pass: All four claims quote the abstract accurately. The engineering implication is explicitly conditional ('If the proposed mapping holds'), and the draft states that implementation and test results are not established. The question follows from the stated falsifiable prediction. The limitations correctly note that only the abstract was available.

From this we derive the falsifiable prediction that selectionist consolidation will outperform instructionist alternatives as memory stores grow large.

media — pass: The selected media is the verified arXiv abstract page, which is the direct original source. No better source media was found.