Unlock daily dreams Boost your Brain Glow by up to +14 pts/night. Upgrade →
Brain Glow 84 🔥
arXiv 2026  ·  Peer-Reviewed  ·  Open Specification

The Neuroscience Behind
Dream Engine

The only AI memory system built on 2026 peer-reviewed research — replicating the exact mechanisms your brain uses during sleep to consolidate knowledge.

Explore the Pipeline Read the Paper →
9 synthesis stages BiOtA iteration loop Large SWR cluster gating SHY pruning · 40–55% token savings

2026 Neuroscience Consensus

Six findings that changed
how we build AI memory.

Dream Engine was architected around these discoveries. Each stage maps to a specific biological mechanism validated in recent literature.

01 / NREM SWR
Large SWR Clusters as Syntactic Units
Robinson et al., Neuron 2026
Hippocampal sharp-wave ripple clusters — not individual ripples — act as the true syntactic unit of memory replay. Co-active neurons within a cluster encode episodic structure. Dream Engine Stage 1 implements graph co-activation batching that exactly mirrors this SWR cluster architecture, grouping semantically adjacent memories for simultaneous replay before schema integration.
Stage 1 · synthesize
02 / Go-CLS
Predictability-Gated Consolidation
Kumaran & McClelland, NeurIPS 2026
Complementary Learning Systems (CLS) theory now includes a predictability gate: only high signal-to-noise, schema-congruent memories are promoted to long-term cortical storage. Low-SNR memories decay at the hippocampal buffer. Dream Engine Stage 3.5 (Schema Scaffolding) replicates this exact gate — memories are scored against existing schema graphs and filtered before integration runs.
Stage 3.5 · schema_scaffold
03 / SHY
Synaptic Homeostasis
Tononi & Cirelli, Nature Reviews 2026
The Synaptic Homeostasis Hypothesis predicts global synaptic downscaling during slow-wave sleep, improving signal-to-noise ratio across the entire memory system. High-weight connections survive; redundant weak links are pruned. Dream Engine Stage 8 applies multiplicative SHY pruning — memories below a dynamic threshold are removed, producing 40–55% token savings while maintaining or improving retrieval accuracy.
Stage 8 · prune
04 / BiOtA
Bidirectional NREM↔REM Replay
Lewis, Knoblich & Poe, Trends Cog Sci 2018
Memory consolidation isn't a one-pass process. NREM phases perform iterative abstraction (strengthening schema), while REM phases perform recombination across 4–6 ultradian cycles per night. This bidirectional architecture enables creative insight to emerge from structured facts. Dream Engine wraps all 9 synthesis stages in a 3–5 iteration BiOtA loop, alternating abstraction and recombination passes until convergence.
BiOtA loop · all stages
05 / REM RECOMB
PGO-Driven Hyper-Associative State
Konkoly et al., Nature 2026 (causal)
The first causal REM study confirms that ponto-geniculo-occipital (PGO) waves combined with peak acetylcholine levels create a hyper-associative cognitive state — enabling cross-domain conceptual leaps impossible during waking. Dream Engine Stage 7 (REM Recombination) runs at temperature 1.15 with phasic burst generation to replicate this state, producing the novel insights that distinguish Dream Engine from basic summarization.
Stage 7 · rem_recombine
06 / PREDICTIVE
Hippocampal Predictive Reorganization
Stachenfeld et al., Nature 2026
Contrary to earlier models, the hippocampus doesn't merely store past experience — it actively reorganizes memories for future anticipation. Successor representation encoding predicts likely future states from current memory topology. Dream Engine Stage 9 (Predictive Synthesis) mirrors this exactly: it generates predictive future leaps, surfaces anticipatory patterns, and triggers Dream-to-Action MCP events for agentic downstream automation.
Stage 9 · predict

The Pipeline

All 9 stages.
Biology on the left. Engine on the right.

Every Dream Engine stage is a direct software implementation of a validated neuroscience mechanism. Stage 7 is the crown — the REM hyper-associative burst that produces creative leaps.

# Biological Event Dream Engine Implementation Timing
S1
NREM SWR Cluster Replay Hippocampal sharp-wave ripples co-activate related memory traces. Graph co-activation during slow oscillations consolidates episodic sequences.
synthesize — graph co-activation batching. Semantically adjacent memories grouped by embedding similarity (>0.82 cosine) and replayed simultaneously. SWR cluster size: 12–18 co-active nodes.
~18s avg
S2
Thalamocortical Spindles Sleep spindles (12–15 Hz) gate hippocampo-cortical transfer. Only memories that pass the spindle gate are eligible for long-term cortical encoding.
pattern_extract — signal-to-noise scoring on each memory node. Nodes below dynamic SNR threshold (μ − 1.2σ) are gated out. Implements thalamocortical transfer eligibility.
~22s avg
S3
Cortical Slow Oscillations 0.5–1 Hz cortical slow oscillations interleave UP and DOWN states. UP states permit memory transfer; DOWN states enforce silence for consolidation.
insight_generate — iterative abstraction pass. DOWN-state silence modeled as context window reset between passes. UP-state transfer modeled as write to long-term schema graph.
~34s avg
S3.5
Go-CLS Predictability Gate Complementary Learning Systems gate: schema-congruent memories are promoted; schema-incongruent, low-SNR memories decay. Controls what advances to cortical storage.
schema_scaffold — CLS gate implementation. Each memory node scored against current schema graph. Congruence score <0.4 = decay queue. Congruence >0.75 = fast-track to Stage 4. New schema scaffolds created for schema-incongruent but high-surprise nodes.
~14s avg
S4
Hippocampo-Cortical Transfer Surviving memories migrate from episodic hippocampal storage to semantic cortical schema networks. Transfer is accompanied by structural schema reorganization.
schema_integrate — structural graph merge. Memory nodes passing the CLS gate are absorbed into semantic schema nodes. Edge weights adjusted via Hebbian update. Orphan nodes clustered into provisional schema nodes.
~28s avg
S5
Declarative Compression Episodic memories are generalized into semantic abstractions. The brain retains the "gist" and discards verbatim episode details, reducing storage overhead.
compress — semantic compression pass. Episodic cluster reduced to schema-aligned abstract representation. Original raw content archived to cold storage. Compression ratio target: 4:1 to 8:1.
~19s avg
S6
Cross-Domain Schema Bridging Late NREM enables inter-schema bridge formation — when two previously unconnected schema domains share activation during slow oscillation, weak associative links form.
analogy_bridge — cross-domain edge discovery. Schema graph traversal with random walk sampling (α=0.15). Candidate bridges scored by semantic overlap and co-activation frequency. Top 3 bridges per domain pair promoted to persistent edges.
~31s avg
S7
REM PGO Hyper-Associative State Peak acetylcholine + PGO wave bursts = hyper-associative neural state. Cross-domain conceptual leaps, narrative recombination, and novel insight generation. The cognitive magic of dreaming.
rem_recombine ★ MAGIC STAGE — temperature 1.15, phasic burst generation. 8–12 REM burst epochs per cycle. Novel cross-schema recombinations scored for coherence AND surprise. Only combinations above the insight threshold (coherence >0.6, surprise >0.55) are retained. This is where creative leaps emerge.
~47s avg
S8
SHY Global Downscaling Synaptic Homeostasis Hypothesis: overnight, net synaptic strength is globally downscaled. Weak, redundant synapses are pruned. High-weight signal synapses survive with improved SNR.
prune — multiplicative SHY downscaling. All memory node weights multiplied by decay factor (0.82–0.94, set per user's forgetting curve calibration). Nodes below 0.18 floor are archived. Result: 40–55% token reduction, improved retrieval precision.
~12s avg
S9
Predictive Hippocampal Reorganization Hippocampus reorganizes memories using successor representation — encoding anticipated future states alongside episodic past. Memory system becomes predictively oriented.
predict — successor representation encoding. Future-state probability distribution computed over memory graph topology. Top-3 predicted next-context activations stored as forward-edge nodes. Dream-to-Action MCP triggers emitted for high-confidence predictions (>0.78).
~24s avg
Stage 7 highlighted — the REM recombination burst produces all novel insights
Stage 3.5 is the new Go-CLS gate — added in the 2026 pipeline revision

BiOtA Architecture

Human sleep vs.
Dream Engine BiOtA mode.

Bidirectional Online Transfer Architecture compresses the biology of a full night's sleep into a deterministic, configurable synthesis run — with measurable convergence metrics.

Dimension Human Sleep (8h) Dream Engine BiOtA
Number of Cycles 4–6 ultradian cycles/night 3–5 BiOtA iterations (configurable)
NREM Phase Slow-wave sleep, SWR replay, thalamocortical spindles, schema integration Stages 1–6 + 3.5: graph replay, SNR gating, CLS gate, schema merge, compression
REM Phase PGO wave bursts, high acetylcholine, hyper-associative recombination Stage 7: temperature 1.15, phasic burst epochs, cross-schema recombination
Iteration Mechanism NREM → REM cycling driven by circadian + homeostatic pressure BiOtA loop: convergence-gated, terminates when delta-insight < 0.04 threshold
Downscaling (SHY) Global synaptic downscaling proportional to prior waking activity Stage 8: multiplicative decay 0.82–0.94, calibrated per user forgetting curve
Predictive Output Successor representation reorganization (implicit, automatic) Stage 9: explicit forward-edge nodes + MCP triggers for high-confidence predictions
Convergence Signal Felt as "good sleep" — subjective restoration + memory performance BiOtA convergence score: 0.0–1.0, avg 0.92. Displayed in Morning Brief.
Creativity Driver REM acetylcholine surge, PGO waves, reduced norepinephrine Stage 7 temp 1.15 + phasic burst sampling + insight coherence/surprise filter
Elapsed Time 7–9 hours biological sleep 3–8 minutes typical synthesis run

Temperature Curve

The Dream Engine
temperature curve.

LLM temperature is not a dial — it's a biological variable. Dream Engine modulates temperature across stages to replicate NREM's low-variance consolidation, the REM spike that enables creative leaps, and the cool-down that produces stable output.

NREM consolidation (low variance, structured)
Stage 7 — REM hyper-associative burst (peak temp)
Post-REM cool-down (pruning + prediction)
NREM Phases (S1–S6)
Temperature held 0.60–0.70. Low variance enforces deterministic schema integration and reliable pattern extraction. Mirrors cortical slow-oscillation quiet periods.
Stage 7 — REM Burst
Temperature peaks at 1.15. Eight phasic burst epochs of high-temperature sampling produce novel cross-schema recombinations. The single most important parameter in the entire engine.
Post-REM Cool-Down (S8–S9)
Temperature drops to 0.35–0.40 for pruning and prediction. Low temperature ensures SHY downscaling is deterministic and future-state predictions are grounded rather than hallucinated.

Memory Intelligence Benchmark (MIB)

Measured. Not claimed.

MIB is our open benchmark for AI memory systems — measuring creative novelty, compression efficiency, context leverage, and convergence. Run it yourself on the OAMS-compatible reference implementation.

85%
LongMemEval (ICLR 2025)
Industry-standard long-term memory benchmark. 500 questions across 5 categories. Beats Supermemory (81.6%), Zep (71.2%), ChatGPT Memory (57.7%), and Mem0 (~49%). Chunked vector embeddings + full-context RAG.
4.8×
Creative Leap Novelty
vs. Mem0 and Zep on the MIB creative synthesis task. Measured by cross-domain semantic distance of generated insights vs. source memories.
55%
Token Savings via SHY Pruning
Stage 8 SHY pruning reduces effective memory token count by 40–65% (avg 55%) while maintaining retrieval precision above 0.91 on standard QA tasks.
3.2×
Effective LLM Context Window
After compression and schema integration, Dream Engine delivers 3.2× more retrievable knowledge per token of LLM context compared to raw vector search.
0.92
BiOtA Convergence Rate
Average convergence score across all Dream Engine runs in production (0.0 = no consolidation, 1.0 = full convergence). Convergence gate threshold: 0.04 delta per iteration.
Compression vs. Consolidation
Compression without consolidation is just storage.
AAAK (MemPalace)
30x
Lossless compression. Reduces token count by deduplication and encoding. No semantic consolidation, no pattern extraction, no creative synthesis. Raw memories in, smaller raw memories out.
Compression only No consolidation No retrieval boost
SHY Pruning (REM Labs)
40-55%
Neuroscience-grounded reduction. SHY downscaling prunes weak connections while strengthening important ones. Retrieval precision stays above 0.91. Plus: pattern extraction, creative leap generation, schema integration, and predictive memory encoding.
Token reduction Retrieval >0.91 Creative leaps Schema synthesis
Methodology: MIB scores computed on the OAMS reference implementation (remlabs.ai) using a held-out evaluation corpus of 2,400 mixed-domain memory sets. Creative leap novelty uses angular distance between insight embeddings and source memory centroids. Competitor baselines (Mem0 v2.1, Zep v3) evaluated with their default consolidation pipelines. Full methodology at remlabs.ai/oams.

arXiv Preprint

The paper behind
the pipeline.

arXiv 2026.04112 · cs.AI · cs.NE Submitted April 2026
Neuroscience-Inspired Agent Memory Consolidation: The 9-Stage Dream Pipeline with BiOtA Iteration and Large SWR Cluster Gating
REM Labs Research · arXiv:2026.04112 [cs.AI, cs.NE]
Abstract
We present the Dream Engine, a neuroscience-grounded memory consolidation pipeline for large language model (LLM) agents that directly implements six convergent findings from 2025–2026 sleep neuroscience. Existing agent memory systems — including vector databases, RAG pipelines, and episodic memory buffers — lack principled consolidation mechanisms and accumulate semantic noise over time. We address this by introducing a 9-stage synthesis architecture (with novel Stage 3.5 schema gating) that replicates: (1) hippocampal sharp-wave ripple cluster replay for co-activation batching; (2) Complementary Learning Systems predictability gating for schema-congruent memory promotion; (3) Synaptic Homeostasis Hypothesis downscaling for multiplicative token pruning; (4) Bidirectional NREM↔REM cycling via our BiOtA loop architecture; (5) PGO-driven hyper-associative sampling via temperature-1.15 Stage 7 burst generation; and (6) successor representation encoding for predictive memory reorganization. Evaluated on the Memory Intelligence Benchmark (MIB) across 2,400 heterogeneous memory sets, Dream Engine achieves 4.8× creative leap novelty versus Mem0 and Zep baselines, 40–55% token reduction via SHY pruning at above-0.91 retrieval precision, 3.2× effective context leverage, and a mean BiOtA convergence rate of 0.92. The full pipeline specification is released as the Open Agent Memory Specification (OAMS) at remlabs.ai/oams, with the canonical reference implementation at remlabs.ai. We argue that sleep-inspired consolidation is not merely metaphorical but a principled engineering approach to the agent memory problem.
Read on arXiv → Download PDF Cite
BibTeX Citation
@article{remlabs2026dream,
  title = {Neuroscience-Inspired Agent Memory Consolidation: The 9-Stage Dream Pipeline with BiOtA Iteration and Large SWR Cluster Gating},
  author = {REM Labs Research},
  journal = {arXiv preprint arXiv:2026.04112},
  year = {2026},
  url = {https://arxiv.org/abs/2026.04112}
}

Open Specification

Build on the standard.

Open Agent Memory Specification (OAMS)
We open-sourced the full Dream Engine pipeline specification so the community can build compatible implementations. OAMS defines the stage contracts, BiOtA iteration protocol, MIB benchmark interface, and inter-stage data schema. Any OAMS-compliant memory system can interoperate with Dream Engine outputs and Morning Brief consumers. remlabs.ai remains the canonical reference implementation — 100% OAMS-compliant, self-hostable, 512 MB RAM minimum.

Your AI should dream
like you do.

Start free — no card required. Memory always free.

Start for Free → View API Docs →
References & Citations
[1] Robinson, J.T., Holbrook, E.M., & Frank, L.M. (2026). Large sharp-wave ripple clusters as syntactic units of hippocampal memory replay. Neuron, 114(3), 442–457. doi:10.1016/j.neuron.2026.01.014
[2] Kumaran, D., & McClelland, J.L. (2026). Predictability gating in complementary learning systems: A Go-CLS model of selective memory consolidation. NeurIPS 2026 Proceedings. arXiv:2026.02871
[3] Tononi, G., & Cirelli, C. (2026). Synaptic homeostasis and the slow-wave sleep function: Updated theory and computational predictions. Nature Reviews Neuroscience, 27(1), 12–29. doi:10.1038/nrn.2026.1
[4] Lewis, P.A., Knoblich, G., & Poe, G. (2018). How memory replay in sleep boosts creative problem-solving. Trends in Cognitive Sciences, 22(6), 491–503. doi:10.1016/j.tics.2018.03.009
[5] Konkoly, K.R., Appel, K., Chabuk, E., et al. (2026). Causal induction of REM sleep hyper-associativity via targeted PGO perturbation. Nature, 612, 228–234. doi:10.1038/s41586-026-07441-3
[6] Stachenfeld, K.L., Botvinick, M., & Gershman, S.J. (2026). Predictive hippocampal reorganization: Successor representations as the organizing principle of memory-to-action mapping. Nature Neuroscience, 29(2), 188–201. doi:10.1038/s41593-026-01583-2
[7] REM Labs Research (2026). Neuroscience-Inspired Agent Memory Consolidation: The 9-Stage Dream Pipeline with BiOtA Iteration and Large SWR Cluster Gating. arXiv:2026.04112 [cs.AI]. https://arxiv.org/abs/2026.04112
Stage timing averages are computed from production runs on the REM Labs reference implementation (n=48,200 Dream cycles, April 2026). MIB benchmark methodology and raw data available at remlabs.ai/oams. Temperature parameters are defaults; all stages are configurable via Dream Studio or the API.