Neuroscience Research

2026 Research Consensus

The neuroscience foundation

Six published findings form the theoretical bedrock of the Dream Engine. Each one maps directly to a stage, mechanism, or parameter in our system.

Stage 1 Basis

Sharp Wave Ripples (SWR)

Robinson et al., Neuron 2026: Large SWR clusters drive memory reactivation in NREM Stage 1. We simulate co-activation batching — grouping semantically related memories for joint reprocessing, directly mirroring hippocampal SWR cluster dynamics.

Robinson et al. — Neuron, 2026

Bidirectional Architecture

Bidirectional Replay

Lewis et al., Trends Cogn Sci 2018: Memory consolidation requires forward NREM replay and reverse REM replay in iterative cycles. This is the theoretical origin of BiOtA — our entire pipeline is a computational instantiation of this bidirectional loop.

Lewis et al. — Trends Cogn Sci, 2018

Stage 3.5 Gate

Go-CLS Gating

Kumaran et al., Neuron 2016: Schema-congruent information transfers preferentially to neocortex via a competitive learning system. We implement this as a congruence filter, blocking low-relevance memories from progressing to schema synthesis.

Kumaran et al. — Neuron, 2016

Stage 8 Basis

Synaptic Homeostasis (SHY)

Tononi & Cirelli, Nat Rev Neurosci 2014: Slow-wave sleep downscales synaptic weights 40–55%, preventing saturation and consolidating signal-to-noise ratio. Our Stage 8 SHY pruning directly replicates this: token reduction averages 52.1% per session.

Tononi & Cirelli — Nat Rev Neurosci, 2014

Stage 7 Mechanism

PGO Waves / REM Theta

Hobson et al.: REM sleep features ponto-geniculo-occipital spike bursts combined with theta oscillations, enabling novel recombination at elevated cognitive "temperature." This directly motivates Stage 7's temperature of 1.15 — the highest in the pipeline.

Hobson et al. — PNAS, 1998

Namespace Architecture

Memory Indexing Theory

Teyler & DiScenna 1986 plus McClelland et al.: Hippocampus as fast-learning index, neocortex as slow-learning storage. Our namespace architecture directly reflects this two-speed system — rapid memory intake with nightly slow-learning consolidation to schemas.

McClelland et al. — Psych Rev, 1995

The BiOtA Pipeline

The 9-Stage Dream Engine

Each stage maps one-to-one to a neuroscience mechanism. Temperature modulates generativity: lower for consolidation, elevated for creative recombination.

Stage	Name	Neuroscience Basis	Temp	What We Do
1	Reactivation	Large SWR clusters (Robinson 2026)	0.70	Graph co-activation batching
2	Pattern Detection	NREM SO-spindle coupling	0.65	Jaccard similarity clustering
3	Schema Transfer	Hippocampal → neocortex transfer	0.70	Schema-congruent synthesis
3.5	Go-CLS Gating	Kumaran et al. 2016	0.75	Filter low-congruence memories
4	Contradiction Resolution	Cortical conflict detection	0.60	Semantic contradiction pruning
5	Priority Scoring	Salience-based reactivation	0.65	TMR priority weights
6	Abstraction	Neocortical schema formation	0.70	Cross-domain synthesis
7	REM Recombination ★	PGO waves + theta oscillations	1.15	Creative leap generation
8	SHY Downscaling	Tononi SHY hypothesis	0.35	Token pruning 40–55%
9	Consolidated Output	Memory-schema integration	0.40	Final memory synthesis

★ Stage 7 is THE MAGIC — the only phase operating above base LLM temperature. Inspired by PGO-wave-driven novel recombination in biological REM.

Computational Neuroscience

BiOtA vs Human Sleep

A side-by-side mapping of biological sleep architecture to our computational implementation. The correspondence is not approximate — it is structural.

Aspect	Human Sleep	BiOtA Engine
Cycles	4–6 ultradian cycles/night	3–5 configurable iterations
NREM Phase	Slow-wave sleep, spindles, SWRs	Stages 1–6 (temp 0.60–0.75)
REM Phase	PGO waves, theta, dream state	Stage 7 (temp 1.15)
Consolidation	Hippocampal → neocortex	SWR clusters → schema synthesis
Creativity	Novel recombination in REM	Cross-domain leaps at temp 1.15
Pruning	SHY downscaling in SWS	40–55% token reduction Stage 8
Convergence	Night-to-night schema overlap	Cosine similarity > 0.85 threshold
Output	Consolidated long-term memory	Insights + creative_leaps + predictions

Temperature Architecture

The pipeline temperature curve

Temperature controls generativity at each stage. NREM-equivalent stages run cool for precise consolidation. Stage 7 spikes to 1.15 — the REM moment where creativity happens.

0.70

Stage 1
Reactivation

0.65

Stage 2
Pattern

0.70

Stage 3
Schema

0.75

Stage 3.5
Go-CLS

0.60

Stage 4
Contradict.

0.65

Stage 5
Priority

0.70

Stage 6
Abstraction

1.15 ★

Stage 7
REM ★

0.35

Stage 8
SHY Prune

0.40

Stage 9
Output

NREM-equivalent (0.60–0.75)

REM-equivalent Stage 7 (1.15)

SHY consolidation (0.35–0.40)

MIB v2 Benchmark Results

Measured performance

The Memory Intelligence Benchmark (MIB v2) measures retrieval quality, compression efficiency, and emergent insight generation across 100 randomized runs on mixed datasets.

biota-bench — MIB v2 — 100 runs

$ biota-bench --runs=100 --dataset=mixed

Benchmark: BiOtA Memory Intelligence Benchmark (MIB v2)

════════════════════════════════════════════════════

Metric Score vs Baseline

────────────────────────────────────────────────────

Retrieval Accuracy94.2% +31% vs vector DB

Novel Insight Generation4.3/session +∞ (baseline: 0)

Context Compression58.4% 3.5x effective context

Convergence (avg)0.987 —

Creative Leaps / Session3.8 —

SHY Pruning Efficiency52.1% —

REM Score (population avg)73 —

arXiv Preprint

The BiOtA paper

The theoretical and empirical basis for our architecture, written for academic review. Cite us if you build on our work.

arXiv — cs.AI / cs.NE — 2026

BiOtA: Bidirectional Iterative Replay for Artificial Memory Consolidation

REM Labs Research Team

Abstract

We present BiOtA (Broader Information Overlap to Abstract), a computational implementation of bidirectional memory replay inspired by Lewis et al. (2018). BiOtA simulates human ultradian sleep cycles across AI memory stores, producing emergent insights through 9-stage temperature-modulated processing. The system implements Sharp Wave Ripple-inspired co-activation batching (Robinson et al., 2026), Go-CLS schema gating (Kumaran et al., 2016), and Synaptic Homeostasis Hypothesis-derived pruning (Tononi & Cirelli, 2014). Stage 7 — the REM-equivalent phase — operates at temperature 1.15, enabling cross-domain creative leap generation. Across 100 benchmark runs (MIB v2), BiOtA achieves 94.2% retrieval accuracy, 4.3 novel insights per session, and 52.1% context compression versus naive vector retrieval baselines. We release all weights, pipeline code, and benchmark tooling as open source.

BibTeX

@article{remlabs2026biota, title = {BiOtA: Bidirectional Iterative Replay for Artificial Memory Consolidation}, author = {REM Labs Research Team}, journal = {arXiv preprint arXiv:2026.XXXXX}, year = {2026}, note = {cs.AI, cs.NE}, url = {https://remlabs.ai/research} }

Built on 30 Years of Sleep Neuroscience

The neuroscience foundation

The 9-Stage Dream Engine

BiOtA vs Human Sleep

The pipeline temperature curve

Measured performance

The BiOtA paper

Science you can run tonight.