← Project overview

Reviewed public research · October 31, 2026

NewBrain memory retrieval: measured evidence and context-limit failures

Executed evidence-ID retrieval only: 30 independent synthetic questions, four existing source routes and five distractor-history scales; no model answers, production memories, neural learning or protected Original12 tests were used.

Measured memory selection

A fresh authored history contains 30 initial facts, 10 linked updates and 10 unrelated records across 50 sessions. Independently frozen questions ask about unchanged facts (10), updated current facts (10), multi-record connections (5) and earlier states (5). The same questions were evaluated at +0, +10, +50, +100 and +200 unrelated sessions, scoring evidence-ID contracts rather than final answers.

Route+0+10+50+100+200
NewBrain Memory7225/3025/300/30*0/30*0/30*
NewBrain Memory7725/3025/300/30*0/30*0/30*
Kira scoped route24/3024/3024/3024/3024/30
Kira legacy route14/3014/3013/3013/3013/30

*At +50 and beyond the NewBrain route selected more current records than the unchanged 64-record admission and failed all 30 with explicit resource insufficiency. Its five smaller-scale misses needed earlier superseded versions. The scoped Kira route missed five historical cases and one multi-record case; the legacy route returned forbidden stale versions for 15 questions.

NewBrain's other fixed resource bound was 32,768 memory bytes; Kira's fixed returned-evidence limit was five. Full NewBrain build_question/codec and answer generation, real conversations, multi-person privacy, ingestion cost and full downstream recall were not tested. An opt-in selection candidate is source-ready, but has not passed new independent data. Reviewed public baseline ↗ · Status metadata ↗

Separate real continuing-learning diagnostics

Three genuinely executed, independently authored continuing practice runs completed 1, 128 and 128 training updates in one native model lineage. All changed parameters and passed exact persistence, but none produced successful complete known/held-out output or retention. The third used an optional parameter-block clipping policy supported by 10 software checks and independent derivative tests. Chronologically corrected teacher-forced loss decreased from 7.48 to 2.136 across 128 updates and its first target became the top-ranked choice (probability 0.903), but complete free generation still failed.

A fourth continuing diagnostic has a chronological record of 572 completed updates with steadily declining teacher-sequence loss. Separate unfamiliar prompts have not passed, and these diagnostics do not establish general language understanding or checkpoint adoption. The accepted protected Original12 remains 2/12 and was not rerun. No private lessons, model arrays, keys, raw answers or operation captures are published. Reviewed learning status ↗