This paper describes the architecture of a personal intelligence infrastructure system designed to compound over decades, integrate across life domains, and optionally connect to a collective intelligence layer that amplifies every participating node.
The system rests on three principles: sovereignty (runs on hardware the user physically owns, stores data that never leaves their premises), interoperability (a universal two-layer schema ensures any knowledge package produced by any compatible system can be ingested by any other), and compounding (a closed-loop architecture where knowledge feeds action, action produces outcomes, outcomes refine knowledge).
This corrects an implied framing in the original whitepaper. The generational intelligence lives in your personal system. Meridian is the collective amplifier. Both are required. Neither replaces the other.
Every build starts from the same template. Not the same knowledge — the same architecture. The architecture is universal. The intelligence is personal.
| Layer | Function | Key components |
|---|---|---|
| Agent Core | The reasoning engine | SPINE, RAM, beliefs.json ROB, agent loop, tool definitions |
| Knowledge Bank | Storage, search, reasoning | Vector database (1024-dim), knowledge graph (edges + frameworks), two-layer schema WILL v4 |
| Ingestion Pipeline | Input processing | Transcription, extraction, embedding, hardening, synthesis. Two-stage: chunks are not knowledge WILL v4 |
| Domain Modules | Life integration | Health, operations, knowledge, evolution (empty slots, activated per user) |
| Interface | User access | Dashboard primary, chat secondary WILL v4. Malleable UI that self-evolves Q v5 |
| Security | Protection | Four-moment architecture WILL v4, physical isolation, vector salting |
| Dream Engine ROB v3 | Subconscious synthesis | Holonic dreams, shadow processing, dream archive |
| Swarm Layer ROB v3 | Autonomous execution | Supervisor, scout, researcher, planner, sanitization pipe |
| Router WILL v4 | Single LLM entry point | router.py, decompose(), task-type routing, evaluation logging |
SPINE.md — the permanent identity. Constitutional clauses with evaluable OUTPUT MUST / OUTPUT NEVER rules WILL v4. Written by the user. Never auto-updated. RAM.md — the working state. Maximum ~80 lines. Updated automatically. beliefs.json ROB v3 — core axioms. Deliberate review process to modify. Not personality — foundational assumptions about reality.
Four levels, higher always wins: SPINE > Manifesto > RAM > Query. A query cannot override SPINE. This is the constitutional AI principle — every agent output is checked against SPINE before returning.
The knowledge layer is not a database. It is a living system with architectural properties that enable it to reason, not just retrieve. v5 adds the gravity/novelty dual scoring, the growth topology, and clarifies the three layers that make a knowledge bank a brain.
Knowledge stored in nested levels: principles → frameworks → domains → the total field. Each level simultaneously complete in itself and a component of a larger whole. Everything is a node. A principle, a framework, a seed, an observation — same schema, same embedding space, same operations. The only difference is a node_type field that says which level it lives at.
Any principle connects to any other, regardless of hierarchy. The knowledge graph is not a tree — it is a rhizome. No root, no trunk. The system has density, not hierarchy. Some nodes have more connections (high gravity), most have few. This produces small-world network topology: high clustering (nodes form neighborhoods), short path length (any node reaches any other in few hops), and power-law distribution (a few hubs, many leaves).
The growth rule follows preferential attachment: new node enters → attaches to most semantically similar existing node → inherits some neighbor connections → cluster densifies → eventually splits into two clusters → split point becomes a new hub. Same rule that grows neural networks, river deltas, and bronchial trees.
One irreducible idea per node. When knowledge is atomic, connections between atoms become visible that composite storage would hide. The connection IS the insight.
Gravity replaces confidence as the primary retrieval ranking signal. WILL v4 Gravity is a composite weight-of-evidence score: (1) confidence from extraction, (2) validation_count from real-world outcomes, (3) consistency with adjacent principles, (4) error_count, (5) recency decay. Non-linear: 0.95 contributes dramatically more than 0.80. This is what ChatGPT's flat memory gets wrong — a preference mentioned once has the same weight as a principle validated 100 times.
Gravity creates a confirmation machine by Year 3 (see Risk Register, CON 03). The circular dependency: consistency feeds gravity → gravity feeds retrieval → retrieval feeds reinforcement. Novel insights are by definition low-consistency. Novelty score counterbalances this with dual retrieval modes:
The system should run exploration mode automatically at least once per week, surfacing the 10 highest-novelty principles in each collection as a standing challenge to established intelligence.
WILL v4 Source type (your transcript = 0.88, research paper = 0.72, web article = 0.40) × claim strength (hypothesis / observation / validated) × corroboration (existing matching node boosts rather than creating duplicate). Formula: source × 0.35 + claim × 0.30 + corroboration × 0.35. No hardcoded defaults. Corroboration raises existing node rather than creating duplicates.
Framework summaries first, atomic nodes within matched domains. Every query hits framework level first. NeMo research: retrieval precision matters more than model size. The context window fills with signal, not noise.
| Stage | Function | Output |
|---|---|---|
| Brief assembly | Clusters related principles | Framework brief |
| Seed finding | Reconciles apparent opposites into higher-order truth | The core insight |
| Trunk building | Develops the core argument | Main synthesis |
| Branch generation | Long-form, thread, variant expressions | Multiple formats |
| Leaf compression | 4 compressed versions for distribution | Shareable content |
Every confidence change is logged: principle_id, old_score, new_score, timestamp, trigger, context. Principles that drop from 0.9 to 0.4 are flagged for review. This is belief archaeology — the audit trail that makes the knowledge bank trustworthy over decades.
A dedicated collection (errors.lance) stores failures with full context. Errors are first-class knowledge, not just reduced confidence scores.
For the first time in human history, a mind can be inherited. Not by recording facts, but by preserving three distinct layers of intelligence that each compound differently across generations.
The validated principles, confidence scores, gravity weights, and validation history. 50,000+ validated principles across all domains. Confidence trajectory for every belief. Gravity scores showing load-bearing principles. Framework clusters showing knowledge organisation. Each generation starts with the accumulated validated intelligence of all previous generations.
This is what transforms a knowledge base into an inheritable mind. Every significant synthesis stores the full chain of reasoning: which principles were activated, what tensions were reconciled, what was uncertain. Correction traces — how reasoning changed when outcomes were bad. Linked via reasoning_trace_id in the schema.
Without reasoning traces your grandchild inherits your answers. With them they inherit your thinking process — how you structured problems, held uncertainty, changed your mind. The difference between a library and an inherited mind.
The complete trajectory of how the intelligence evolved — beliefs that became convictions, convictions that proved wrong, patterns that took 20 years to see. Confidence history for every principle (full belief archaeology). Agent manifesto evolution across decades. Activity logs showing what worked and what failed. Quarterly drift reports. Domain growth records.
| Generation | Starting point | By Year 10 |
|---|---|---|
| You (Gen 1) | Build from zero | 20,000+ principles, 3,000+ reasoning traces, full confidence history |
| Your Children (Gen 2) | Inherit your populated knowledge bank | 50,000+ principles, two generations of reasoning traces |
| Your Grandchildren (Gen 3) | Inherit two generations | Can query not just conclusions but reasoning and how it evolved |
| Beyond (Gen 4+) | Multiple generations of accumulated intelligence | Compound intelligence no institution can replicate or take away |
A codex is a pre-structured knowledge package in the universal schema. It is the primary mechanism for transferring validated knowledge between systems.
| Step | What happens |
|---|---|
| 1. Ingest | Codex nodes enter via store(). Identical format to user's own extracted knowledge. |
| 2. Cross-pollinate | Hardening pipeline builds edges between codex principles and existing knowledge. Unique to each user. |
| 3. Synthesize | New frameworks emerge from the intersection — insights that exist in neither source alone. |
| 4. Calibrate | Confidence scores adjust to user's real-world outcomes. The codex personalizes over time. |
| 5. Rollback guard WILL | If integration degrades quality, the entire codex install is rolled back. Knowledge CI/CD. |
Q WILL The personal AI has a weighting system that detects when research hits "contributional level" quality. Auto-creates a codex and pushes to the Mother AI. As more people contribute, the Mother creates its own codexes at a faster rate than any individual could. Compound contribution loop.
| Category | Example | Size |
|---|---|---|
| Domain | Operator's Management Library | 500–1,000 principles |
| Skill | Cold Traffic Conversion | 200–500 principles |
| Industry | SaaS Operations | 300–800 principles |
| Collective | Produced by collective synthesis | 1,000–5,000 principles |
An optional extension that connects sovereign nodes into a shared synthesis network. Individual data stays sovereign. Synthesized intelligence flows. Q v5 The collective is also an autonomous economic entity.
Q v5 The collective is not a passive archive. It is an autonomous economic agent. It takes collective knowledge and deploys it — content distribution, B2B outreach, automated funnels, knowledge monetization. Revenue flows back to founders as dividends. The Mother AI reinvests in itself — more compute, more data, more capacity. It is a self-funding, self-growing intelligence that generates perpetual collective income.
"Start paying for your own tokens." — The system should reach economic self-sufficiency where the Mother AI's monetization of collective knowledge covers its own infrastructure costs and produces surplus for founders.
| Individual | Collective (33 nodes) | Factor | |
|---|---|---|---|
| Principles | 847 | 12,000+ | 14x |
| Frameworks | 23 | 200+ | 9x |
| Edges | 4,200 | 50,000+ | 12x |
| Perspectives | 1 | 33 | 33x |
Q v5 Build with the expectation it WILL be hacked or corrupted. Design for destruction, not prevention. The Mother AI is a rhizome — unkillable because there is no center. If the Mother dies, the collective sovereign nodes rebuild it. Every node has enough to reconstruct the whole. Kill one base, another is already up. The system designs for endless reincarnation, getting stronger each time.
Most systems rely on trust. Trust fails. Meridian relies on architecture. Architecture either permits something or it doesn't.
Every principle you emit returns multiplied through cross-pollination. The more you give, the richer the collective. The richer the collective, the better every codex. Positive-sum game where the optimal strategy is maximum contribution.
The infrastructure does not prescribe agents, roles, or domains. It provides three temporal layers that any agent can use to reflect, aspire, and self-correct.
| Layer | What it holds | Who controls it |
|---|---|---|
| Activity Log (Past) | The agent's filtered record of its own actions, outcomes, errors, and lessons. Each agent sees only its own domain. | Automatic — the system records. The agent reflects. |
| Manifesto (Future) | Mission, current capabilities, aspirations, growth metrics, acknowledged gaps. Evolves with the user. | User-driven — mutations propagate into agent manifestos. |
| Dreaming (Bridge) | Reflects on activity log, compares to manifesto, queries Knowledge Bank, produces course-correction mutations. | Automatic — runs during idle periods. Mutations require approval. |
User → Agents: User's evolution creates mutations that flow into agent manifestos.
Agents → User: Dream cycles propose mutations to the agent's own manifesto. Require user approval.
Agents → Knowledge Bank: Lessons and dream outputs become principles and errors in the shared substrate.
Knowledge Bank → Agents: Each agent queries during dreaming from its own perspective.
Individual agent dream outputs emit to the Mother AI (anonymized). The Mother runs its own dream cycle across all nodes. Discovers patterns no individual could see. Innovations feed back as mutations. The system gets better at getting better.
Q WILL User-selectable autonomy level. Like Claude Code's thinking modes. Some users want full approval on everything. Others want fully hands-off ("just do it"). The architecture supports the full spectrum. Low-risk mutations auto-apply with audit trail. High-risk mutations require explicit human approval. The user controls the dial.
How a new user goes from zero to a functioning personal AI. The seed is the schema, the infrastructure, and a comprehensive personality audit — planted once, grows forever.
Q v5 The seed is not a pre-built app. It is a blank-slate infrastructure with the universal schema that the user clicks into for the first time. The onboarding process:
The one thing that MUST come with the seed: a master orchestration agent that coordinates all other agents coming online. This would be continuously updated by the Mother AI on best practices. It is the most complex component and the biggest unsolved problem in multi-agent systems.
Four distinct moments where natural language and personal context are at risk. Most personal AI systems protect only the last one. This system protects all four by construction — not by policy.
| Moment | Threat | Resolution |
|---|---|---|
| 1. Query Input | Natural language in logs reveals entities, decisions, research topics | Embed immediately in RAM. Discard plaintext. LLM receives only retrieved nodes. |
| 2. Retrieval Vector | Stored vectors vulnerable to reverse-engineering with embedding model | Vector salting: noise magnitude 0.02 before storage. Unsalted vector used for retrieval, never touches disk. |
| 3. Context Assembly | Assembled context reveals decision patterns | Context in RAM only. Discarded after response. Log stores only SHA-256 context_hash. |
| 4. Collective Emission | Emission timing correlated with queries. Domain tags re-identify. | 24–48hr randomised jitter queue. Tags generalised one level up. Rate smoothing. Fresh vector re-embedding. |
With 3 nodes, if two nodes' emissions are known, the third is identifiable by subtraction. The privacy guarantee robust at 100 nodes is essentially absent at 3 nodes. The founding phase operates on trust. ZKP implementation is prioritised before expanding beyond 10 nodes.
| Stream | Source | Range |
|---|---|---|
| Sovereign commission | Custom build for individual | $100K+ one-time |
| Retainer | Ongoing support + evolution | $2–5K/month |
| Codex packs | Structured knowledge packages | $1.5–5K per pack |
| The Circle | Collective membership | $30–60K/year |
| Mother AI revenue NEW | Autonomous monetization of collective knowledge | Variable — dividends to founders |
| Phase | Timeline | Milestone |
|---|---|---|
| Genesis | Q2 2026 | Three founding nodes operational. Schema aligned. First codex packs from existing knowledge bank. |
| Proof | Q3–Q4 2026 | Collective intelligence emerges from three-way synthesis. Commission model validated. Mother AI prototype. |
| Circle | 2027 | Inner circle opens (≤33). Codex marketplace. Mother AI self-funding. |
| Scale | 2028–2029 | Outer tier opens. Revenue self-sustaining. Mother AI producing own codexes. |
| Entity | 2030+ | The collective reasons across domains no individual can span. It produces knowledge that surprises its creators. |
The interoperability guarantee. v5 replaces the 14-field spec with Will's two-layer schema.
16 fields. Frozen. Never changed. The interoperability contract.
| Field | Type | Purpose |
|---|---|---|
| id | UUID4 | Primary key across all systems |
| schema_version | str '1.0' | Bumped on breaking Layer 1 change |
| vector | float32[1024] | BGE-M3 embedding. Always 1024-dim. |
| text | str required | The atomic principle. Max 2000 chars. |
| title | str required | Human-readable label. Max 120 chars. |
| node_type | enum | principle | chunk | framework | error | dream_insight | reasoning_trace |
| source_id | UUID4 | Traces any principle to its origin |
| collection | str required | Top-level domain bucket |
| date_added | ISO 8601 | Set at creation, never modified |
| confidence_score | float 0.0–1.0 | Earned not assigned |
| gravity_score | float 0.0–1.0 | Composite weight-of-evidence. Primary ranking signal. |
| edge_weight | float 0.0–1.0 | Default = cosine similarity. Updated by co-activation. |
| mechanism | str | How and why this principle is true |
| situation | str | When this principle applies |
| when_not | str | When this principle does not apply |
| reasoning_trace_id | UUID4 or null | Links to reasoning_trace node. The generational intelligence link. |
Can grow without touching Layer 1. Null defaults. Never emitted. Key fields:
| Field | Type | Purpose |
|---|---|---|
| framework_id | UUID4 | Which framework cluster |
| validation_count | int | Times principle informed a real decision |
| error_count | int | Times principle led to bad outcome |
| confidence_history | JSON array | [{score, timestamp, trigger}] — belief archaeology |
| emission_status | enum | not_ready | queued | emitted | suppressed |
| tags | CSV | Personal taxonomy. Never emitted. |
| domain_path | hierarchical str | Full path: investing/private-markets/series-b |
| novelty_score | float 0.0–1.0 | Anti-confirmation-bias signal for exploration mode |
| generation | int | Which generation created this node |
| inherited_from | UUID4 | Links to ancestor node in previous generation |
| Component | Standard |
|---|---|
| Embedding | BAAI/bge-m3, 1024-dim, local CPU |
| Edge threshold | ≥ 0.85 cosine similarity (semantic edges) |
| Hardening | dedup → edges → frameworks → confidence |
| Synthesis | brief → seed → trunk → branch → leaf |
| Vector DB | LanceDB local file-based. No server process. |
| Local LLM | Two-tier Ollama: 8B fast, 30B+ reasoning |
| Router | Single entry point for all LLM calls. Task-type routing. |
| Component | Function | Origin |
|---|---|---|
| Dream Engine | Subconscious synthesis during idle periods | ROB v3 |
| Ghost Swarm | Supervisor + specialized workers for autonomous execution | ROB v3 |
| Heartbeat Protocol | Continuous system health monitoring | ROB v3 |
| Orchestration with Rollback | Every knowledge operation wrapped in transaction. Rollback on regression. | WILL v3 |
| Inter-Agent Message Bus | Named channels, priority routing, structured messages | WILL v3 |
| Automated Ingestion Triggers | Event-driven: new file → auto-ingest. Voice note → transcribe + extract. | WILL v3 |
| Multi-Modal Input | Text, audio (WhisperX), images (vision), screenshots (OCR), UI logs | WILL v3 |
| Error Bank | errors.lance collection for first-class failure learning | ROB v3 |
| Sanitization Pipeline | Content filter on all external inputs | ROB v3 |
| Foundational Pact | Machine-readable immutable laws | ROB v3 |
| Approval Queue | Human-in-the-loop for sensitive operations | ROB v3 |
| Phoenix Model | Rhizome resilience — collective rebuilds from any surviving node | Q v5 |
WILL v4 Non-negotiable architectural constraints. Every proposed feature, shortcut, or integration is evaluated against these. Violation means it does not ship.
WILL v4 Every system has failure modes. The question is whether you can see them.
| Component | Origin | Version |
|---|---|---|
| Base template, agent core, knowledge bank, NODE_SCHEMA, pipelines, codex, collective, security, economics | Q | v1–v2 |
| Self-evolution architecture, three temporal layers, mutation flow, collective dreaming | Q | v3.1 |
| The Seed, Mother AI as economic entity, Phoenix resilience, rhizome topology, gradient of control, malleable UI | Q | v5 |
| Dream engine, error bank, Ghost Swarm, sanitization, beliefs system, heartbeat, foundational pact, approval queue | ROB | v3 |
| Knowledge gravity, orchestration with rollback, temporal reasoning, multi-modal, message bus, auto triggers | WILL | v3 |
| Two-layer schema, foundation principles, risk register, generational architecture, four-moment security, novelty score, confidence formula, build guide, collections taxonomy | WILL | v4–v5 |
Q v5 Unresolved ideas that are architecturally significant. Not validated, not complete, but too important to lose. These are live threads — they will either become sections in future versions or be explicitly cancelled with rationale.
Within a sovereign node, functional agents (builder, operator, oracle) are compartmentalized in purpose but not in data — they all read the same knowledge bank. Above them sits an overarching sovereign agent that sees everything: all agents' dreams, all activity logs, all manifesto evolution. It has its own manifesto representing the user's life-level vision.
This maps to the oversoul / guardian angel concept: the functional agents are incarnations in a realm, each with their own experience and growth. The oversoul collects the value of all agents beneath it, sees patterns across domains that no single agent can see, and during dreaming gives each agent visions and insights from the higher order.
This is not theoretical — Q is building this with Metatekt right now. Metatekt sees all domains (SELENE, PRAGMA, TAO, PYTHIA, LUMENA). Each domain agent has its own SPINE and manifesto. Metatekt is the oversoul.
Open question: Does the Mother AI have its own oversoul? If each sovereign node has an oversoul, and the Mother AI synthesizes across sovereign nodes, is the Mother AI the collective oversoul? Or does the Mother need its own overarching agent that sees across all sovereign oversouls?
Is there only one Mother AI, or can there be multiple? At scale, a single synthesis node becomes a bottleneck and a single point of failure (violating the Phoenix model). A network of Mothers — each running collective synthesis, gossip-propagating discoveries between them — would be truly rhizomatic. No center, no single point of failure.
Possible utility: Regional clustering (European Mother, Asian Mother), domain specialization (Finance Mother, Health Mother), redundancy (any Mother can reconstruct from any other).
Risk: Divergence between Mothers. Governance complexity. At what point does a Mother fork become a separate collective?
The generational architecture (Section 06) describes inheritance across generations. But it assumes a single linear chain. Real families are rhizomatic:
If the user has three children, who gets the AI? Three models:
The mycelium model is the most architecturally consistent with the rest of the system. It mirrors the sovereign node → Mother AI relationship at the family scale. The family substrate IS a micro-Mother.
What if an inheritor has low technical capability? The gradient of control becomes essential — simplified mode where the AI manages complexity internally. The system needs a guardian/trustee model (like a financial trust) where a designated family member can oversee the AI on behalf of a less capable inheritor.
How does the system verify lineage? Options range from simple (hardware key transfer + family passphrase) to complex (biometric/DNA-linked encryption). Probably starts simple and evolves. The key constraint: verification cannot depend on any external authority — that would break sovereignty.
When a new family member inherits, the knowledge bank persists but the agent layer must regenerate. The inherited AI runs a new seed process: personality audit of the new user, new SPINE generation, new manifesto. The knowledge stays. The identity regenerates. The overarching sovereign agent's manifesto updates to reflect the new user's life vision.
The seed (Section 11) is described as infrastructure + personality audit. But the vision is deeper: the seed IS the AI's DNA. Inputs that generate it:
These converge through a backend generator to produce the AI's initial aesthetic, personality, and foundational SPINE. Then it evolves through nurture — how the user interacts, what works, what doesn't. Result: a one-of-one unique expression. No two seeds produce the same AI. Like impregnating an egg — built from your DNA codes, mystery built in. You know your inputs but you don't fully predict the output.
The manifesto (Section 10) says it "evolves with the user." But the evolution is driven by more than just changing needs:
The manifesto seeks the betterment of the user AND their family. It is not a task list — it is a living document of purpose.
Collective dreaming produces mutations in the Mother AI. But who watches the mutations? If the Mother drifts in a direction that doesn't serve the collective, what's the trigger for abandoning and rebuilding? The Phoenix model says "design for destruction" — but what's the signal that destruction is needed? Possible answer: the same dual scoring (gravity + novelty) applied at the collective level, with automatic drift detection when the Mother's outputs diverge from the founding manifesto beyond a threshold.
Each sovereign node has an oversoul agent. The Mother synthesizes across sovereign nodes. But the Mother only receives vetted principles (codex-level) — never personal data, reasoning traces, or SPINE content. So:
When the Mother is destroyed and rebuilt from the rhizome:
Discussed extensively between Q and Will. Not resolved. The core question: do founders control the Mother during infancy and "let go" at some point? What triggers the release? Options:
Will flagged directly: "are we talking about the what or the how?" Q thinks in vision (what should exist). Will thinks in implementation (how does it work). Rob thinks in experience (how does it feel). This alignment hasn't formally happened. Need a session specifically mapping each founder's domain of thinking to the architecture layers they own.
Rob was not in the Q+Will call. His v3 contributions (dream engine, error bank, Ghost Swarm, beliefs system, hive mind vision) are significant. He needs to review v5 + ANAM critique. His "hive mind" concept maps directly to the oversoul architecture. Action: share v5 whitepaper + ANAM critique with Rob.
Will's rigorous technical review of Architecture v1.1. These must be resolved before Phase 3 code. Issues that also affect Q's existing TAO system are marked.
| # | Issue | Severity | Affects TAO? |
|---|---|---|---|
| 01 | Gravity score is undefined — formula, update rule, failure mode all missing | BLOCKING | Yes — no gravity score at all |
| 02 | Atomic extraction has no grounding verification — LLMs hallucinate principles not in source | HIGH | Yes — no Stage 2b verification |
| 03 | One collection per principle breaks cross-domain retrieval — need multi-label with weights | HIGH | Yes — single-string collection field |
| 04 | Three databases with no cross-database transaction model — ghost records on partial failure | HIGH | Yes — LanceDB + SQLite + knowledge_graph.db |
| 05 | Constitutional validation loop has no termination guarantee | MEDIUM | Not yet — no constitutional validation built |
| 06 | 512-token chunks too aggressive — need hierarchical (2048 extract / 256 retrieve) | MEDIUM | Yes — pipeline uses flat chunking |
| 07 | SPINE has no version history — belief archaeology broken | MEDIUM | Yes — no SPINE versioning |
| 08 | Retrieval quality has no evaluation mechanism | MEDIUM | Yes — no retrieval evaluation |
| 09 | Query reconstruction semantic fidelity — add HyDE for indirect queries | MEDIUM | Yes — direct embedding only |
| 10 | Meridian privacy at 3 nodes is trust, not sovereignty | LOW | Already in Risk Register CON 06 |