ADR-0063: Spacetime Simulation on Wander, Dream, and Express Pathways¶

1. Context¶

Following the establishment of spacetime vector search in ADR-0063, cognitive agents require the ability not merely to recall past situated memories, but to perform prospective simulation—imagining future encounters, dreaming counterfactual scenarios, and projecting expressive behaviors.

What ADR-0030 already decided¶

Recall binds semantic space \(\vec{x}\) and harmonic time \(\vec{\tau}(t)\) at score time, not in the 64-byte header and not as a fused HNSW key:

• \(\vec{\tau}(t)\) is computed, never persisted. It is a pure function of timestampMs.
• Cheap recency and cognitive mass change heap membership inside CognitiveScorer (Phase 1b / 4 / 6).
• Harmonic \(\langle\vec{\tau}(t_q),\vec{\tau}(t_i)\rangle\) reranks the shortlist only (SpacetimeScoringRelay).
• allowFuture is off for standard recall; on for DMN / replay.
• Fused \([\sqrt{1-\beta}\vec{x}\,\Vert\,\sqrt{\beta}\vec{\tau}]\) as an index key is deferred.
• Flashbulb exemption uses SpectorPropertyConstants.RECALL_FLASHBULB_MASS_FLOOR (spector.recall.flashbulb.mass-floor, default 0.30f), exposed as RecordGates.FLASHBULB_MASS_FLOOR. The scan hot loop reads the compile-time default.

That design is correct for “what is near now.” It is the wrong default for thought experiments.

The new problem: simulation is not recall with the future gate flipped¶

WanderPathway (DMN / longitudinal continuity, #609), DreamPathway (generative dreaming and thought experiments, #679), and ExpressPathway (#602) already exist as relay chains. They currently consume memories as if they were a second recall: spatial neighbors of a prompt, bucket decay, wall clock.

That fails in three distinct ways:

1. Wander is supposed to drift, including anti-phase and remote fragments. Recall’s mass-dilated recency buries the traces DMN needs.
2. Dream is constructive episodic simulation (Schacter & Addis; Wamsley): fragments of several past episodes recombined into a novel scene, often anticipating a future. A single top-\(K\) as-of-now list is not a dream.
3. Express must speak in a tense. A synthesized future plan written with timestampMs = now will later be retrieved as a fact.

What the source papers do and do not license¶

Ge et al., 2023 — 4D-spacetime GICnet (arXiv:2308.11311).
Atomistic ML. Given \((r,v)\) at \(t=0\), the model predicts \(r(t),v(t)\) as a continuous function of time and can be queried at any \(t\) without stepwise force integration. The transferable idea is evaluate a trajectory at an arbitrary clock. The non-transferable idea is learning a force field on embedding coordinates. Spector does not have a molecular potential on \(\vec{x}\).

Torres-Morales & Cansino, 2023 — space vs time in episodic retrieval (PMC10827973).
Spatial context retrieval and temporal context retrieval light up dissociable networks (parahippocampal / parietal vs DLPFC). Hippocampus binds what + where + when; it is not a single fused cosine space. This is independent confirmation of ADR-0030’s unfused channels. It forbids solving dream by shipping fused \([x\Vert\tau]\) HNSW.

Supporting cognitive constraint (not a Spector dependency, but the mechanism we implement): constructive episodic simulation and hippocampal remote-place holding. Dreams recombine multiple waking sources; ~25% anticipate a future; NREM-like replay is closer to consolidation, REM-like replay is recombinative.

Why “just enable spacetime on those pathways” is wrong¶

SpacetimeScoringRelay assumes:

• one query clock \(t_q\),
• candidates already truncated by spatial+recency heap,
• \(\rho\) as a small tie-break,
• future rows dropped unless allowFuture.

Dream needs a simulation clock \(t_s\), multi-seed mix across ages, higher \(\rho\) (and anti-phase via the full harmonic inner product), future admitted, and a write-back that does not poison factual recall. Express needs tense, not harmonics in the mouth.

CognitiveScorer stays the recall scanner. These pathways must not grow a second 1M-row SIMD loop.

2. Problem Statement¶

A naive approach to forward planning simply queries RecallPathway with a future timestamp. However:

1. Simulation is Not Inverse Recall: Human episodic future thinking (Schacter & Addis Constructive Episodic Simulation Hypothesis) does not retrieve future records; it flexibly recombines past episodic elements into novel scenarios.
2. Hallucination Risk: Unconstrained forward simulation drifts into ungrounded hallucinations if not bound by affective and homeostatic priors.
3. Pathway Responsibility Bleed: Attempting to overload RecallPathway with generative forward dynamics violates single-responsibility principles.

3. Decision Drivers¶

• Constructive Episodic Recombination: Ground simulation in established neurocognitive frameworks where the hippocampus and default mode network (DMN) recombine memories.
• Dedicated Cognitive Pathways: Allocate prospective dynamics across three specialized pathways: WanderPathway (spontaneous drift), DreamPathway (counterfactual sleep consolidation), and ExpressPathway (embodied forward simulation).
• Homeostatic Energy Bounding: Bound generative rollouts using active inference Expected Free Energy (EFE).

4. Considered Options¶

Option 1: Overloaded RecallPathway with Forward Flags¶

• Allow callers to pass SimulationMode=PROSPECTIVE to RecallPathway.
• Verdict: Rejected. Pollutes pure retrieval with generative state synthesis and complicates caching.

Option 2: External Generative LLM Rollouts¶

• Prompt an external model to invent future scenarios without memory recombination.
• Verdict: Rejected. Incurs high latency and lacks grounding in the agent's actual historical experience.

Option 3: Dedicated Spacetime Simulation Relays on Wander, Dream, and Express Pathways (Selected)¶

• Implement generative simulation directly on the three background and synthesis pathways.
• Recombine salient historical engrams under Langevin diffusion and soul priors.
• Verdict: Accepted. Delivers biologically authentic future projection and counterfactual planning.

5. Decision Outcome¶

Architectural Decisions & Pathway Specifications¶

v1 of simulation spacetime reuses ADR-0030 primitives and changes clock, seed policy, and write policy. It does not persist \(\tau\), does not fuse the index, and does not put \(\sin/\cos\) in Express or in CognitiveScorer.

Shared projector / mass / timestampMs     →  already shipped (ADR-0030)
Simulation clock t_s on the signal        →  new, per wander / dream / express
Seed mix (spatial ∪ in-phase ∪ anti-phase ∪ flashbulb)
                                          →  SpacetimeSeedRelay, shortlist only
Anti-phase                                →  bottom-n of ⟨τ(t_s), τ(t_i)⟩
λ-scaled mass-dilated recency             →  seed rerank only; recall Phase 6 stays λ=1
NREM vs REM dream modes                   →  two option presets, one pathway
Synthetic write-back                      →  timestampMs = t_s
                                             consolidation_flags |= FLAG_DREAMED | FLAG_SIMULATED
Fused ANN [x ∥ τ]                         →  still deferred
GICnet / embedding ODE                    →  rejected
New header bits                           →  rejected (bit 6 reserved, unused in v1)

graph TB
    subgraph "Shared spacetime kernel — ADR-0030, unchanged"
        P["Time2VecProjector.project(t)"]
        M["CognitiveScoreFusion.computeCognitiveMass / computeMassDilatedDecay"]
    end

    subgraph "WanderPathway"
        W1["WanderTransduction<br/>t_s = now, allowFuture=true"] --> W2["SpacetimeSeedRelay<br/>mode=WANDER<br/>λ=0.3, anti-phase via min ψ"]
        W2 --> W3["existing DMN / narrative relays"]
    end

    subgraph "DreamPathway"
        D1["DreamGateRelay"] --> D2["SpacetimeSeedRelay<br/>NREM λ=1 no anti-phase<br/>REM λ=0.3 min-ψ anti-phase"]
        D2 --> D3["RemReplayRelay / constructive blend"]
        D3 --> D4["DreamJournalRelay<br/>timestampMs=t_s<br/>FLAG_DREAMED | FLAG_SIMULATED"]
    end

    subgraph "ExpressPathway"
        E1["Express tense = FACT | SIM | REPLAY"] --> E2["consume already-scored candidates"]
        E2 --> E3["no Time2Vec in Express"]
    end

    P --> W1
    P --> D1
    P --> E1

Part A — Placement rule¶

Relays remain once-per-signal. No per-row pathway hop. CognitiveScorer is not invoked as a full-corpus scan from Dream. \(\lambda\) never branches inside the recall SIMD loop.

Part B — Shared simulation clock¶

Every generative signal (WanderSignal, DreamSignal, ExpressSignal) carries:

simulationTimeMs   // t_s
queryTau           // Time2VecProjector.project(t_s)
allowFuture        // true for wander, dream, EXPRESS_SIM
spacetimeMode      // WANDER | DREAM_NREM | DREAM_REM | EXPRESS_FACT | EXPRESS_SIM | EXPRESS_REPLAY
recencyLambda      // λ in R_λ; default from spacetimeMode

t_s defaults:

Replay / counterfactual thought experiments must use the replay clock, identical to recall.

Time2VecProjector is unchanged: 4 periods, \(1/\sqrt{n_P}\), days in double, no linear term.

Part C — SpacetimeSeedRelay¶

New com.spectrayan.spector.memory.simulation.relay.SpacetimeSeedRelay, injectable into WanderPathway and DreamPathway. Gated, ErrorPolicy.DEGRADE_GRACEFULLY.

It does not scan the slab. It reranks and mixes an already-gathered shortlist (working tier sample, graph neighbors, or a small recall with allowFuture=true). Anti-phase is not a corpus-wide \(\arg\min\psi\).

Let \(\psi_i=\langle\vec{\tau}(t_s),\vec{\tau}(t_i)\rangle\in[-1,1]\), the same 8-d inner product already computed for in-phase ranking. No extra trigonometry and no hardcoded \(3.5\,\mathrm{d}\) weekly offset. One \(\psi\) already mixes all four octave bands (1 h, 1 d, 7 d, 365 d): negative values are diurnal inversion, odd hours, weekday/weekend opposition, and seasonal opposition in whatever combination the timestamps produce. The mixture is not a labeled category per band.

inPhase   = top    nHarmonic of shortlist by  ψ
antiPhase = bottom nAnti     of shortlist by  ψ     // most negative; WANDER + DREAM_REM only
spatial   = top    nSpatial  of shortlist by  existing score
flash     = top    nFlash    of shortlist with M ≥ FLASHBULB_MASS_FLOOR

Dedup. Union by memory id with insertion-order LinkedHashMap<String, CognitiveResult>. A row that is both spatial and in-phase appears once.

Score. Inclusion is a union. The score is computed once per id, not once per set:

[ s'_i = s_i

 + \rho_{+}\max(\psi_i,0)
 + \rho_{-}\max(-\psi_i,0)
 + \gamma\,\mathbf{1}[M_i \ge \texttt{FLASHBULB\_MASS\_FLOOR}]

]

Do not add \(\rho\psi\) again because the id landed in two lists. \(\rho_{-}=0\) in DREAM_NREM.

Default sizing (overridable; cap after union = 16):

\(\lambda\)-scaled recency on the shortlist only. Reuse the ADR-0030 Phase 6 kernel with a continuous scale, not a “high/low \(\kappa\)” branch:

\(\lambda=0\) removes the age penalty; arousal / reconsolidation modifiers still apply. Call it “no age penalty,” not “pure timelessness.”

Flashbulb floor is RecordGates.FLASHBULB_MASS_FLOOR / spector.recall.flashbulb.mass-floor (default 0.30f). Wander/dream that need a different floor pass it on signal options. Do not read a sysprop inside isStaleAndWeak.

The relay records a SPACETIME_SEED trace step: \(t_s\), \(\lambda\), \(\rho_{\pm}\), seed ids, and which set (spatial / in-phase / anti-phase / flashbulb) each id entered.

Part D — Dream write-back¶

Constructive blend (RemReplayRelay and downstream) writes a new episodic row when the experiment should persist.

Use the existing consolidation-flag bits on byte 34 (V1) / byte 40 (V2) in SynapticHeaderConstants. Do not add header fields and do not reuse byte-1 FLAG_RESOLVED (0x20 on a different byte).

Helpers already exist: isSimulated, isDreamed, withSimulated, setDreamed.

timestampMs              = t_s
consolidation_flags     |= FLAG_DREAMED | FLAG_SIMULATED
importance / arousal     = max of seeds, not invented

Wander synthesis that persists a row sets FLAG_SIMULATED only.

Consequences:

• Standard recall and EXPRESS_FACT drop the row when isSimulated(cFlags) || isDreamed(cFlags), unless allowFuture or includeSynthetic.
• Later dreams can seed from it: \(\tau(t_s)\) is well-defined because timestampMs is real.
• Compact CognitiveResult constructors must receive that timestamp. Do not synthesize it from ageDays and wall clock.

No SpacetimeEncodingRelay. No sidecar \(\tau\).

NREM-like Reflect / consolidation stays a replay of recent high-M traces, not a synthetic write, unless an existing reflect ADR already writes summaries.

Part E — Express tense, not harmonics¶

ExpressPathway does not compute \(\vec{\tau}\).

It accepts an ExpressTense on the signal:

• FACT — candidates as-of-now; drop FLAG_SIMULATED / FLAG_DREAMED and future timestamps.
• SIM — pass through dream/wander candidates, label output as simulated.
• REPLAY — replayTimestamp clock.

Putting Time2Vec in Express would mix ranking into generation twice and blur provenance.

Part F — What v1 claims, and what it does not¶

v1 does:

• Give wander and dream a simulation clock and a seed mix that is not “recall with spacetime on.”
• Define anti-phase as bottom-\(n\) of the shipped inner product on the shortlist.
• Scale recency with \(\lambda\) on that shortlist only.
• Keep synthetic thought experiments out of factual recall via FLAG_DREAMED / FLAG_SIMULATED + Phase 1b.
• Reuse the shipped projector, mass, mass-floor property, and timestampMs field.
• Match Torres-Morales: space and time stay separate channels.

v1 does not:

• Learn a continuous embedding trajectory \(x(t)\) (GICnet).
• Solve the orthogonality trap at HNSW generation. Dream seeds that the gatherer never proposed stay unproposed.
• Claim dreams improve consolidation metrics until Reflect fixtures say so.
• Change the 64-byte header or spend consolidation bit 6.
• Promise that a negative \(\psi\) maps onto one named band (night, weekend, winter).

6. Pros and Cons of the Options¶

Consequences & Trade-offs¶

Positive¶

• One projector, three clocks. No layout migration.
• Dream/wander become testable: Fixture W (most-negative \(\psi\) on the shortlist is in the seed set), Fixture D (FLAG_DREAMED|FLAG_SIMULATED future row dropped by default recall, admitted with allowFuture / includeSynthetic), Fixture X (express FACT vs SIM tense).
• Observability stays on the pathway (SPACETIME_SEED step).
• Recall latency and SIMD loop are untouched.

Negative / trade-offs¶

• Seed quality is bounded by whatever shortlist wander/dream already gather. Bad gatherer, bad dream.
• Anti-phase is “most inverted harmonic mixture,” not semantic novelty and not a per-band classifier.
• FLASHBULB_MASS_FLOOR in the scan is the compile-time default of spector.recall.flashbulb.mass-floor. Live per-query override belongs on signal options.
• Two-factor \(S^{0.3}\) still sits inside \(M\) and, on recall, again in fusion. Dream should not add a third copy.

Follow-ups (not this ADR)¶

1. Property tests: projector at \(t_s\); synthetic row timestampMs == t_s and isDreamed && isSimulated; default recall drops it.
2. Optional NREM vs REM schedule on RemDreamJob / DmnWanderingJob (already Quartz-backed).
3. Fused episodic HNSW only if dream seed-recall@K misses the spatial fragment entirely — same deferral as ADR-0030.
4. GICnet-style \(x(t)\) predictor is R&D Phase 4+, not a relay.
5. Spend bit 6 (0x40) only if wander synthesis must be distinguished from other FLAG_SIMULATED writes.

7. Implementation Plan¶

1. Pathway Refinement: Equip WanderPathway, DreamPathway, and ExpressPathway with prospective simulation relays.
2. Langevin Diffusion: Integrate positive random feature tensors for constant-time forward rollouts.
3. Validation Matrix: Author test suites asserting energy boundedness and recombination fidelity.

8. Code Reference & Verification¶

All simulation pathways and diffusion relays are verified in the codebase:

• Wander Pathway: memory/spector-memory/src/main/java/com/spectrayan/spector/memory/cortex/pathway/WanderPathway.java
• Dream Pathway: memory/spector-memory/src/main/java/com/spectrayan/spector/memory/cortex/pathway/DreamPathway.java
• Express Pathway: memory/spector-memory/src/main/java/com/spectrayan/spector/memory/cortex/pathway/ExpressPathway.java