Skip to content

DREAMING — the "LLM dreams over context to form novel connections" idea (researched 2026-07-24)

Julian's concept: take accumulated session/memory context, let the model hallucinate/recombine on it (like dreams recombining memories) to form connections that wouldn't otherwise form, discard the useless, keep the novel. Two fanned research passes (AI-dreaming/memory line + computational- creativity line). Verdict: the concept is textbook, but the full mechanism is unbuilt in peer- reviewed form, and the selection/distance half is genuine open ground.

The precise academic name for the idea

Combinational creativity (Boden) realized as Blind-Variation-Selective-Retention (Campbell/ Simonton) over a flat associative hierarchy of remote associations (Mednick). - Boden 1990/1998 — combinational / exploratory / transformational creativity. Yours = combinational (novel combos of familiar ideas), definitionally. - Campbell 1960 (BVSR) — creativity = generate variations blindly + selectively retain the good. This IS "hallucinate widely, then prune." Simonton is its modern formalizer. - Mednick 1962 — creativity = forming remote associations; "flat associative hierarchies" reach distant concepts. This is your "random associations between concepts," from 1962. - Conceptual blending (Fauconnier & Turner 1998/2002; computationally: Goguen, COINVENT, Confalonieri 2018) — the mechanism under recombination. Historically hard to implement for lack of math — the gap LLMs now fill. - LLM-era term already in circulation: "combinatorial creativity" (arXiv:2412.14141).

The dream-specific full mechanism — who states it

  • Gwern, "LLM Daydreaming" (gwern.net/ai-daydreaming, Jul 2025) — states YOUR whole loop verbatim: a background Day-Dreaming Loop, generator explores non-obvious concept-pair links, critic filters for novelty, survivors written back and compound. Frames LLMs as lacking a "default mode network." NOT peer-reviewed (essay). The complete idea is articulated but unbuilt.

Peer-reviewed / credible-lab pieces (each implements a SUBSET)

  • Sleep-time Compute — Letta + UC Berkeley, arXiv:2504.13171 (2025). The exact "model does offline compute on its context while idle" ENGINE. Dual-agent: live agent + background sleep agent reprocessing/reorganizing memory in downtime. Emphasis = precompute/consolidate, NOT novelty+critic.
  • Generative Agents' reflection — Park et al., arXiv:2304.03442 (2023). Periodically synthesizes stored memories into higher-level insights fed back to memory. The synthesis half; no random recombination, no discard step.
  • Cooking Up Creativity — arXiv:2504.20643 (2025, TACL). Structured recombination for novelty via "creative leaps." Recombination half, no offline loop.
  • LLMs can Realize Combinatorial Creativity — arXiv:2412.14141 (2024). Closest single paper to the framing: implements Boden's combinational creativity in an LLM with cross-domain retrieval to force distant combos + structured recombination.
  • A-Mem — arXiv:2502.12110 (2025). Zettelkasten-style: dynamically links memories into a network → "connections that wouldn't otherwise form."
  • Vocabulary roots: Wake-Sleep (Hinton 1995, literal sleep phase generates fantasies), World Models / Dreamer (Ha-Schmidhuber; Hafner, "training in dreams") — representation/control, not conceptual recombination. DeepDream (2015) = feature viz, unrelated.
  • Memory-as-sleep-replay: Deep Generative Replay (Shin et al., NeurIPS 2017) + 2024-25 brain- inspired continual-learning line (arXiv 2504.14727, 2509.00047).

Empirical LLM-idea-generation (your loop, already partly built)

  • "Can LLMs Generate Novel Research Ideas?" — Si, Yang, Hashimoto (Stanford), arXiv:2409.04109 (2024). 100+ researchers: LLM ideas judged MORE novel than experts', weaker on feasibility. Also documents a diversity ceiling (models recycle ideas). The key evidence the generate step works.
  • SciMON (2305.14259), ResearchAgent (2404.07738), The AI Scientist / v2 (2408.06292, 2504.08066) — all = over-generate → novelty-check + LLM-critic filter → (some) run experiments.
  • Hallucination-as-feature line: survey arXiv:2402.06647; "Shakespearean Sparks" 2503.02851; "Does Less Hallucination Mean Less Creativity?" 2512.11509 (reducing hallucination suppresses divergent creativity — direct evidence for the tradeoff). CAUTION: Peeperkorn 2405.00492 — "temperature is a RISK parameter, not a creativity parameter" (cranking randomness ≠ novelty).
  • Analogy roots: Copycat (Hofstadter & Mitchell 1994/95) — generate-associations-then-filter with a literal temperature knob; the purest historical prototype of the whole loop.

THE WHITE SPACE (honest — this is where a contribution lives)

DON'T pitch "recombine + filter" as a new theory — it's the single most-worked idea in computational creativity (1960→2024); the exact experts you want to impress will dismiss an over-claim (see BRAND.md AI-slop/over-claiming warning). The genuinely OPEN problems: 1. The FILTER is the bottleneck, not the generator. Every system generates plenty of "novel" candidates but can't reliably separate good-novel from merely-weird (feasibility lags novelty). A better convergent/selection stage = real white space. 2. Forcing genuinely DISTANT combinations. Diversity ceiling (Si et al.); sampling isn't truly blind (Gabora arXiv:1409.2210); temperature ≠ novelty (Peeperkorn). Mechanisms guaranteeing conceptual distance are underexplored. 3. Blind vs. guided variation is unsettled (Campbell vs. Gabora) — live theoretical dispute. 4. Novelty evaluation itself is immature — ad-hoc LLM/embedding checks; a principled novelty filter is open ground. 5. Julian's specific angle: the continuous background loop over accumulated session/memory context (Gwern's DDL) has NO peer-reviewed, evaluated implementation. Building it = first working version of an essay-only idea.

The keystone opportunity (ties to BRAND.md)

This is a candidate for THE named, buildable, publishable artifact. A defensible project: - Implement the day-dreaming loop over Claude Code session/memory context (the substrate already exists in ~/.claude memory + transcripts). - Contribution NOT "the idea" but: a distance-forcing generator (cross-domain retrieval) + a principled novelty+usefulness critic, evaluated ("does it produce ideas humans rate novel AND useful above a baseline / above raw temperature sampling?"). - Name it. Open-source with the eval. Clean citation lineage: Gwern DDL, Sleep-time Compute, Generative Agents, Boden/Campbell/Mednick, Si et al. - Venue: EMNLP Industry/workshop or an ICCC (computational creativity) track. This is the "artifact chain" keystone from BRAND.md, made concrete.

BUILD PLAN — what it takes to build + open-source (2026-07-24)

The system (~a weekend for MVP; Python + Claude API + embedding model; loop is a few hundred lines)

  1. Corpus it dreams over = accumulated context as a concept store. Natural source: ~/.claude memory files + session transcripts (or any KB). Extract concepts/claims/entities, embed them. A few thousand nodes is plenty.
  2. Distance-forcing sampler (CONTRIBUTION #1). Do NOT sample randomly, do NOT just crank temperature (temperature = risk param, not creativity param — Peeperkorn). Pick concept pairs/ triples that are FAR in embedding space but have a plausible bridge. Controlled conceptual distance is the real novelty lever.
  3. Generator (the "dream"). Prompt Claude for a non-obvious connection/idea between the distant concepts. Generate many, cheaply, in a background idle loop (sleep-time-compute style).
  4. Critic / selective retention (CONTRIBUTION #2, the real bottleneck). Multi-dimensional, not one "is this good?" call: novelty (embedding distance to whole corpus + prior survivors, so it's genuinely new not reworded) × usefulness/feasibility (rubric) × coherence gate (kills distance- forced garbage). The joint novelty-AND-useful frontier is where the field is stuck.
  5. Memory writeback + lineage. Survivors compound and recombine across runs — closes the loop.

The eval (this IS the paper; ~80% of the work)

Question: does the loop produce ideas humans rate novel AND useful, above baselines? Baselines to beat, in priority: (1) raw high-temperature sampling ← beating this is THE headline result; (2) reflection-style synthesis (Generative Agents); (3) random-pair (no distance control). Human panel rates novelty + usefulness (small 3-5 rater study + LLM-judge validated against them). Metrics: joint novelty×usefulness, diversity (escape the recycling ceiling?), survival rate. The rated idea-set is itself a citable dataset artifact. Research risk = LLM-judges noisy on novelty; a defensible novelty metric is part of the contribution.

Effort (honest)

MVP that dreams + prints survivors: a weekend. Publishable version: ~3-5 focused weeks, ~80% eval + writeup not code.

Open-source + publish path (= the BRAND.md keystone, concrete)

Name it (short/greppable) → clean repo (reference impl + one-command demo + eval harness + results table + rated-idea dataset) → canonical human-written blog post on own domain → HF Papers submit → 6-8pg paper → ICCC (International Conference on Computational Creativity, natural home, friendlier bar than NeurIPS) or an EMNLP/NeurIPS creativity workshop. Citation lineage reviewers want: Gwern DDL (the unbuilt proposal being realized), Sleep-time Compute (engine), Generative Agents (reflection), Boden/Campbell/Mednick (theory), Si et al. (empirical anchor). Claim = NOT the idea but the first working, evaluated, distance-controlled implementation with a principled critic, run over real memory. That's the artifact chain: named method → runnable code → benchmark → paper.

Recommended entry point

Build the weekend MVP first — dream over THIS session's memory + transcripts, eyeball the survivors. If they make you go "huh, interesting" → the eval weeks are worth it. If mush → learned cheap, iterate the critic. Proposed home: ~/Desktop/I/dreaming/ (extractor · distance-sampler · dream generator · first-pass critic, pointed at ~/.claude memory as the corpus).

PRIOR-ART SCAN — is it already BUILT? (2026-07-24) → NO mature/evaluated version exists

  • sgoedecke/idea-mill — closest built attempt; real public repo of Gwern's loop, but author calls it "pretty half-assed", semi-manual (hand-curated facts in YAML b/c model wouldn't generate them reliably), DID produce "a few genuinely novel ideas", est. needs ~30% more work. NO rigorous eval, no baselines. A blog prototype.
  • davidrd123/latent-dreamer — 0 stars; implements Mueller's 1990 DAYDREAMER architecture, NOT Gwern's loop; cross-session accumulation "not proven yet"; tests architectural correctness, no novelty metrics.
  • RogueCtrl/OpenClawDreams — OpenClaw plugin, nightly "dream cycle" / surreal narrative memory toy; not a novelty-gen-with-eval system.
  • Academic cousins (Deep Ideation, SPARK, SciMON, AI Scientist) = generator→critic idea loops over SCIENTIFIC LITERATURE, not the "dream over accumulated personal/session memory" angle, none framed as the DDL. VERDICT: the exact scoped artifact (distance-forcing generator + principled novelty+usefulness critic + run over accumulated memory + real eval beating raw-temperature/reflection baselines + named + open-sourced) DOES NOT EXIST. Ideal prior-art situation: validated-enough (idea-mill got real ideas) but nobody did it rigorously → clear citable gap + low bar (beat "half-assed").

ORDER-OF-MAGNITUDE PLAN (2026-07-24) — beyond the competent version

Competent version (fill gap, workshop paper) = linear. OOM = become the REFERENCE POINT. Four levers: 1. Own a VERIFIABLE eval → become the benchmark. Dream over a TIME-FROZEN corpus (pre-2025 only), then measure how many dreamed connections were independently discovered/published in 2025-26. Objective ground truth for novelty+value (no hand-wavy LLM judge panels). Headline number: "dreamed 400 connections, 23 later found independently, raw sampling got 3." The harness = THE benchmark for idea-generation → everyone who measures against it cites you (highest- citation act in ML per BRAND research). 2. Ship a DAEMON, not a demo. Background process pointed at any corpus (Obsidian vault, codebase, Zotero, Claude Code memory) that dreams overnight → morning connections. Two starving audiences: PKM/Obsidian crowd (dead-notes pain) + agent-memory ecosystem (Letta/Mem0/A-Mem all consolidate, none GENERATE). Adoption by a Letta-class framework = upstream dependency (llama.cpp lesson: dependency is the deepest citation). 3. One undeniable dream. Run until it produces ONE connection that leads to something verifiably real (finding/feature/bug), tell it with receipts + logs ("dreamed on date X"). The AI-Scientist moment. 4. Compounding curve over months. Long-horizon run showing survivors seed better dreams — nobody has shown this; cheap (runs while sleeping); difference between technique and system. Unfair advantages: distribution machine already built (LOOPS/X per BRAND.md); Gwern reliably links serious implementations of his proposals → day-one reach. Cost answer to Gwern's "daydreaming tax": cheap-model generator + expensive-model judge cascade (itself a publishable efficiency result). Formula: verifiable eval (benchmark) + daemon (dependency) + one real hit (story) + compounding curve (moat) × existing distribution = own the category.