ADR-0060: Cognitive Continuity Layer and Decoded Mind Streams¶

1. Context¶

This record documents exploratory research into hypothetical brain-computer interface (BCI) sensory streaming, neural decoding, and cognitive continuity telemetry.

0. Classification & Anti-Claims (read first)¶

This document is an internal speculative ladder, not a partnership pitch, medical claim, or product roadmap commitment.

We do NOT claim that Spector or Neuralink (or any public BCI) can today:

• Read or dump autobiographical “brain memories” as biological engrams
• Transfer thoughts or identity between brains
• Write memories back into neural tissue
• Act as an on-implant or FDA device memory layer
• Replace electrophysiology stores (NWB / DAQ lakes)

What this does explore: storing and reactivating decoded cognitive artifacts (symbols, latents, episodes) derived from a BCI (or any high-bandwidth intent channel), using Spector as a software continuity layer — with a far-horizon research ladder toward hypothetical write-back that remains unsolved science.

1. Motivation¶

Project Lead’s framing: if a BCI can surface what a brain is “thinking” (intent / content), can Spector store that state and later transfer it?

Public BCIs (e.g. Neuralink Telepathy / PRIME materials) today primarily decode movement or communication intention into actions (cursor, clicks, typing; speech and vision as future/BDD directions). That is intent I/O, not mind upload.

Still, the strategic question is legitimate as a north star:

Can a cognitive memory engine provide continuity for a stream of decoded mind content — across sessions, devices, agents, and eventually (if science allows) back toward the biological substrate?

Spector’s existing strengths (off-heap SIMD engrams, ACT-R-ish activation/decay, graph/Hebbian association, AISME kernels, import/export) map cleanly to the middle of that pipeline — not to raw spike storage or biological write-back.

9. References (public orientation)¶

• Neuralink Technology / PRIME materials — https://neuralink.com/technology/
• ClinicalTrials.gov PRIME NCT06429735
• Musk & Neuralink 2019 JMIR research platform paper (wired, animal)
• NWB — https://nwb.org/
• Industry motor BCI decode literature (BrainGate lineage) — intent features, not memory engrams

2. Problem Statement¶

During exploratory design sprints, theoretical proposals suggested extending Spector's episodic memory substrate to directly ingest decoded neurological mind streams (fMRI/EEG/neural telemetry):

1. Premature Speculation: Neural decoding technology remains highly experimental, noisy, and non-standardized.
2. Architectural Scope Creep: Implementing bespoke BCI ingestion protocols risks diverting engineering resources away from core agentic memory, active inference, and multimodal conversation.
3. Ethical & Data Privacy Unknowns: Ingesting unconstrained continuous neural signals presents severe unresolved governance and user consent challenges.

3. Decision Drivers¶

• Focus on Production Agentic Memory: Spector must excel as a dependable, production-ready cognitive memory platform for software agents and developers.
• Scientific Rigor: Reject ungrounded claims regarding full brain-state decoding or digital consciousness uploading.
• Honest Status Governance: Formally mark speculative or out-of-scope explorations as Withdrawn rather than leaving them in an ambiguous state.

4. Considered Options¶

Option 1: Implement Dedicated BCI Mind-Stream Module¶

• Build custom ingestion pipelines for continuous neural telemetry.
• Verdict: Rejected and Withdrawn. Out of scope and scientifically premature.

Option 2: Standard Multimodal Episodic Ingestion (Selected)¶

• Restrict Spector's production scope to concrete multimodal streams: conversational text, acoustic prosody, visual embeddings, tool executions, and internal active inference dynamics.
• Verdict: Accepted. Maintains sharp focus on high-impact production capabilities.

5. Decision Outcome¶

2. Conceptual Pipeline¶

flowchart LR
    B["brain"] --> R["BCI read (sparse neural sampling)"]
    R --> D["decoder"]
    D --> I["intent / text / speech hyp. / latent z_t"]
    D --> S["Spector Cognitive Continuity Layer<br>episodes · latents · graph · decay<br>activation · export / import"]
    S --> A["software agent / co-pilot / other Spector store ✅ near-term"]
    S --> F["decoder⁻¹ / stim policy → brain ❓ far-horizon"]

Transfer therefore splits into two meanings:

1. Software transfer — export/import a continuity store between devices/agents. Feasible.
2. Biological transfer — induce the same memory in neural tissue. Not available; do not market.

3. What We Would Actually Store¶

Not “the brain.” Store versioned cognitive artifacts:

Invariant: raw kHz×channel electrophysiology stays in a proper neuro data lake (NWB etc.). Spector holds summaries and decoded products only, unless a future specialized binary layout is explicitly designed for binned features (still not “memories of the brain”).

4. Spector Mapping (near-term)¶

Reuse, don’t reinvent:

• Kernel / engrams — off-heap records for episodes + latents
• Activation / decay / cognitive mass — which continuity traces stay “alive”
• HNSW + BM25 — recall by similarity and lexical content
• Graph — associative walk (“what was I doing when…”)
• AISME — free-energy / surprise / change-point as software segmentation & salience (metaphorical vs biology)
• Import/export (see existing memory import/export ADRs) — software “transfer”
• Synapse — on-device or user-sovereign deployment profile

New thin pieces (if we prototype):

1. DecodedMindEvent schema (JSON/Avro) — no medical device types
2. Adapter: intent/HID/text stream → remember pipeline
3. Continuity export bundle (encrypted, user-keyed)
4. Explicit non-clinical packaging + anti-claim banner in docs

5. Ten-Year Ladder (phases)¶

Phases are gates, not calendar promises. Each gate requires evidence before the next is funded as more than paper RnD.

Phase A — Intent continuity log (0–18 months) · Engineering¶

• Ingest decoded events only (synthetic + public BCI feature datasets, or assistive HID streams).
• Spector stores episodes; user can “resume context.”
• Success: measurable assistive UX gain vs key-value prefs; zero neural PHI in multi-tenant cloud by default (on-device first).
• Exit anti-pattern: any marketing that says “brain memory backup.”

Phase B — Latent mind-stream store (1–3 years) · Research + product-adjacent¶

• Persist decoder latents z_t with model provenance; reactivation = feed latents to software agents (not brains).
• Study stability: does the same “thought neighborhood” reappear under decoder drift?
• Success: latent neighborhoods remain meaningful across decoder versions via adapters; export/import works across machines.
• Blocker: decoder drift, non-identifiability of latents, consent for continuous capture.

Phase C — Cross-agent / cross-device continuity (2–5 years) · Platform¶

• Transfer continuity bundles between Spector instances (user-controlled).
• Optional: “digital twin” agent that speaks/acts with user’s stored prefs + episodic context.
• Success: user recognizes continuity; revocation and selective amnesia work.
• Still not: biological write-back.

Phase D — Hypothetical write-back research (5–10+ years) · Science fiction until proven¶

• Only with academic/clinical partners: can patterned stimulation (or future write BCIs) bias recall or percepts in a controlled task?
• Spector’s role would be policy + content sequencing for stimulation recipes derived from stored latents — never unsupervised “memory injection.”
• Hard blockers: inverse problem (latent → safe stim), specificity of memory circuits, adverse effects, ethics, FDA/device pathway, irreversibility.
• Default posture: do not staff this phase as product engineering; revisit only if peer-reviewed write primitives exist.

6. Threats, Ethics, and Governance¶

Treat continuous decoded streams as more sensitive than chat logs even when not clinically labeled.

10. One-Line Summary¶

Spector can be the continuity disk for decoded mind streams; it cannot yet be — and must not claim to be — the transferable substrate of a biological mind.

6. Pros and Cons of the Options¶

7. What Success Looks Like (honest)¶

Near-term win: Spector becomes the best cognitive continuity store for high-bandwidth human intent channels (BCI decode, AAC, wearable agents) — software transfer of context and identity artifacts.

Long-term dream (unproven): those artifacts become inputs to safe, consented write pathways. That dream does not authorize present-tense claims.

7. Implementation Plan¶

8. Recommended Immediate Actions¶

1. Accept this doc as north-star speculative under Jarvis ownership; status remains Draft.
2. If prototyping: Phase A only — decoded-event schema + on-device Synapse profile + synthetic streams.
3. Partner path: assistive-tech / AAC / BCI-app vendors — not implant OEMs as first call.
4. Do not open a Spector epic titled “Neuralink integration” without a signed research boundary note.

• Disposition: Formally marked as Withdrawn. No code will be merged into spector repositories for decoded neural mind streams.

8. Code Reference & Verification¶

• Absence of BCI Code: Verified that no speculative neural decoding stubs exist in nucleus/, memory/, or synapse/.
• Production Scope Alignment: Production ingestion remains strictly governed by RememberPathway and SpectorMemory.remember().