slopshop.gg / dream-engine / pipeline

Dream Engine Pipeline

9-stage neuroscience-inspired memory synthesis. BiOtA bidirectional replay. Large SWR clusters. Free forever.

v4.0 POST /v1/memory/dream/start POST /v1/memory/dream/biota POST /v1/memory/context-dream Free

Section 01

The Full Pipeline

Each stage mirrors a distinct neurological process. Temperature controls divergence: low = consolidation, high = recombination. Stage 7 intentionally spikes — that is where creative synthesis occurs.

Stage	Name	Biological Analog	Temp	Output
1	Reactivation	Large SWR clusters (Robinson Neuron 2026)	0.70	Co-activated memory clusters
2	Pattern Detection	NREM SO-spindle coupling	0.65	Recurring patterns
3	Synthesis / Transfer	ASC hippocampo-cortical dialog	0.70	Merged schema
3.5	Schema Scaffold gating pass	vmPFC Go-CLS gating	0.75	High-SNR schemas only
4	Contradiction Resolution	Schema updating	0.60	Resolved contradictions
5	Priority Scoring	Intelligence Score computation	0.65	Salience-ranked insights
6	Abstraction	Neocortical schema formation	0.70	Higher-order abstractions
7	REM Recombination ← THE MAGIC	REM theta + PGO waves	1.15 🔥	Creative leaps + what-ifs
8	SHY Downscaling	Synaptic Homeostasis (Tononi)	0.35	40–55% token savings
9	Consolidated Output	Predictive reorganization	0.40	Morning Brief + actions

Stage 3.5 is a gating pass, not a billable stage. Schemas below SNR threshold are dropped before Stage 4.

Section 02

BiOtA Mode — Bidirectional Online Transfer Algorithm

Human sleep runs 4–5 ultradian cycles per night. BiOtA mirrors this: NREM forward pass builds schemas, REM reverse pass recombines them, feedback loop refines convergence. Run 3–5 iterations for maximum insight yield.

biota-config.js
// BiOtA — 3-5 iterations (mirrors human ultradian sleep cycles)
// NREM forward → REM reverse → feedback loop
 
POST /v1/memory/dream/biota
{
  "iterations": 4,
  "swr_cluster_threshold": 0.75,       // min co-activation score
  "convergence_threshold": 0.85,       // stop early if reached
  "namespace": "default"
}
 
// Response
{
  "biota_id": "biota_4f9a",
  "iterations_run": 4,
  "convergence_score": 0.92,
  "nrem_schemas": 18,
  "rem_recombinations": 7,
  "tokens_saved": 58,                    // % via SHY downscaling
  "intelligence_score": 91.3
}

iteration anatomy
Each iteration: Stage 1–6 (NREM consolidation, forward) → Stage 7 (REM recombination, reverse) → Stage 8–9 (SHY pruning + output). Convergence is measured as cosine similarity between successive consolidated memory states. Once ≥ convergence_threshold, the loop terminates early.

Section 03

ContextDream — Token Compression for LLMs

Pass any large context window through the Dream pipeline. Returns a semantically compressed gist, creative leaps surface from Stage 7, and a predictive slice sized to your token budget. Effective context 3.2× the raw window.

context-dream.js
POST /v1/memory/context-dream
{
  "context": "Your long LLM context here...",
  "token_budget": 32000
}
 
// Response
{
  "gist": "...",                         // NREM-consolidated summary
  "creative_leaps": ["...", "..."],      // Stage 7 REM outputs
  "predictive_slice": "...",             // fits token_budget exactly
  "tokens_saved": 19840,                 // absolute token reduction
  "rem_score": 0.88                      // recombination quality (0-1)
}

Section 04

Temperature Curve

Temperature controls the model's divergence at each stage. Stages 1–6 consolidate (0.60–0.75). Stage 7 spikes hard to 1.15 — this is where novel associations form. Stages 8–9 cool sharply to lock in signal and prune noise.

0.70

0.65

0.70

S3.5

0.75

0.60

0.65

0.70

1.15 REM PEAK 🔥

0.35

0.40

Bars proportional to temperature. S8–S9 cool to lock signal; noise pruned by SHY downscaling.

Section 05

Benchmark Results

MIB v2 (Memory Intelligence Benchmark). Baselines: Mem0 (Series A, managed), Zep (OSS). Scores averaged across 200 memory sessions with 500–5,000 memories each.

benchmark — MIB v2

Dream Engine vs Baselines (MIB v2):

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Creative Leap Novelty: 4.8× vs Mem0/Zep

Token Savings (SHY): 55–65% via pruning

Effective Context: 3.2× raw window

BiOtA Convergence: 0.92 avg (4 iters)

Insight Quality (MIB): 4.1× actionable

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Section 06

Quick Start

Three calls get you the full pipeline. All endpoints accept JSON. Authenticate with Authorization: Bearer $SLOP_KEY.

Start a Dream

bash
# Kick off all 9 strategies
curl -X POST /v1/memory/dream/start \
  -H "Authorization: Bearer $SLOP_KEY" \
  -d '{"strategies": ["synthesize","pattern_extract","insight_generate",
                  "compress","associate","validate","evolve","forecast","reflect"]}'
 
// → { "dream_id": "drm_7f2b", "stages": 9, "status": "running" }

Run BiOtA Mode (full neuroscience pipeline)

bash
curl -X POST /v1/memory/dream/biota \
  -H "Authorization: Bearer $SLOP_KEY" \
  -d '{"iterations": 4}'
 
// → { "biota_id": "biota_4f9a", "convergence_score": 0.92, ... }

Get your Context Pack

bash
curl -X POST /v1/memory/context-dream \
  -H "Authorization: Bearer $SLOP_KEY" \
  -d '{"context": "...", "token_budget": 32000}'
 
// → { "gist": "...", "creative_leaps": [...], "tokens_saved": 19840 }

CLI shorthand
slop dream run --strategy synthesize,compress,insight_generate
slop dream run --biota --iterations 4
slop context-dream --budget 32k < context.txt