Reviewed public research · November 1, 2026 · Source committed November 1 UTC
NewBrain: bounded historical context selection
Qualified memory-only evidence selection, not an answer-generating model. The independently checked opt-in candidate has not been integrated into the production conversation runtime. The protected Original12 benchmark remains 2/12, not rerun.
800 executed retrieval calls
Separate authors froze 250 distinct-session synthetic observations and 40 questions; 37 were answerable and three unknown. A blind auditor derived answers independently before labels were consulted. Three unknown-wording cues were corrected and re-audited before the first retrieval call. Four actual isolated memory-source routes (Memory72 and Memory77, each default and opt-in), five nested noise-history scales and 40 questions produced 800 retrieval calls. An independent postrun review recomputed 800/800 cells and 60 aggregate groups, checked frozen source/input/result hashes and observed no after-the-fact tuning. The dataset is consumed.
| Isolated route (each backend) | +0 | +10 | +50 | +100 | +200 |
|---|---|---|---|---|---|
| Memory72/77 current default | 23/37 | 23/37 | 0/37* | 0/37* | 0/37* |
| Memory72/77 opt-in selector | 34/37 | 34/37 | 34/37 | 34/37 | 34/37 |
*At +50/+100/+200 the defaults failed native context admission for all 40 questions, including the unknowns. They were not neural answer mistakes. The opt-in selector met unchanged 64-record and 32,768-byte serialized context caps, with no forbidden evidence.
At the smallest scale, current questions stayed 23/26; historical questions changed from 0/11 with current-only eligibility to 11/11 with opt-in historical selection. This adds a temporal retrieval capability, not a clean ranking-only ablation. Three remaining misses require a prior version excluded by eligibility. The three unknowns earned no vacuous credit.
What this does not prove
The evaluation compiled native memory-only append, eligibility, attributed-message and selector definitions under closed AST allowlists. It intentionally excluded learner imports. Full native build_question/codec framing, generated answers, training, retention, real conversations, production integration, multi-owner privacy and protected Original12 testing were not executed. The single-owner corpus lacks harder shared-name distractors and genuine two-link cases. The recorded latency is not an end-to-end speed benchmark. The original selector and opt-in wrappers remain a separately versioned candidate, not an adopted production route.
No raw history, teacher materials, model weights, questions, private identities or private research assets are included in this public summary.
Separate continuing-learning result
In a different genuine native-model study, the fourth prospective practice stage completed its fixed cap of 1,024 neural updates and 30 generations. Final save/reload and source integrity succeeded, but all complete known, held-out and retention outputs failed. Lower teacher-sequence loss did not transfer to complete unfamiliar answers. A fifth, larger curriculum of 64 lessons and a scheduled 2,048 updates/332 generations is ongoing and unscored; scheduled operations are not completed results. No repaired language-learning claim follows, and Original12 remains 2/12.