Skip to main content

Cumulative Consciences

Cumulative Fragmented Consciences & On‑Demand Reasoning​

This document extends the Brainframe architecture by describing how Fragmented Consciences evolve over time, how on‑demand reasoning workflows create new knowledge, and how that knowledge is stored, indexed, and reintegrated into the brain.

This creates a system where every new question makes the system smarter.

1. Overview

In Part I, Brainframe established:

  • Context buckets
  • Fragmented consciences
  • Pre‑reasoned knowledge
  • Goo Balls
  • The final LLM step
  • The ability to request new reasoning when needed

Part II explains how the system grows, learns, and self‑improves:

Every on‑demand reasoning task produces new structural knowledge that can be saved, indexed, and used to power future reasoning.

This transforms a stateless LLM into a continuously growing organisational intelligence.

2. Two Equal Pillars: Pre‑Reasoning & On‑Demand Reasoning

2.1 Pre‑Reasoned Knowledge​

These are scheduled workflows:

  • nightly
  • hourly
  • on data change
  • by bucket

Pre‑reasoning gives:

  • summaries
  • deltas
  • hypotheses
  • cross‑bucket stable knowledge
  • trend detection
  • risk scanning
  • embedding indexes

This forms the “predictable baseline mind”.

2.2 On‑Demand Reasoning (Triggered by LLM Tool Calls)​

When the final LLM step discovers that:

  • the pre‑reasoned material is insufficient,
  • or a new perspective is required,
  • or multiple buckets need a fresh recombination,

…it calls a tool like:

run_bucket_reasoning

This executes:

  • multi‑step reasoning loops
  • custom perspectives
  • cross‑bucket logic
  • targeted pipelines
  • new insights that were not precomputed

The result:
New reasoning, born inside the query.

3. Making the Brain Cumulative

Every on‑demand reasoning output is valuable.
It represents a perspective that did not previously exist.

Brainframe should:

  1. Store every on‑the‑fly reasoning result
  2. Index and categorise it
  3. Embed it in vector search
  4. Attach it to the relevant buckets
  5. Mark it with metadata:
    • time
    • perspective used
    • buckets involved
    • confidence
    • granularity
    • relevance scores
  6. Expose it to future queries

This turns every question into:

  • more intelligence
  • more perspectives
  • more refined summaries
  • more ways to see the organisation

Result:​

The brain is not static.
It grows with use.

4. The Full Learning Loop

When a user asks a question:

  1. Goo Ball is assembled
  2. Final LLM evaluates pre‑reasoned knowledge
  3. If insufficient, it calls run_bucket_reasoning
  4. Worker orchestrator:
    • fetches bucket data
    • runs pipelines
    • calls secondary LLM(s)
    • generates new reasoning
  5. Final LLM uses that to answer the question
  6. Brainframe stores the new reasoning as:
    • a new fragment node
    • vector embedding
    • structured metadata
  7. A “Propagation Engine” optionally sends this reasoning to:
    • other buckets
    • long‑term summaries
    • upcoming nightly cycles
    • anomaly detectors
    • policy generators

This creates cumulative intelligence.

5. Categories of Stored Reasoning Outputs

Brainframe can classify each saved reasoning output into:

5.1 Micro-fragments​

  • A single insight
  • One hypothesis
  • A small correlation
  • Clarification of one factor

5.2 Meso-fragments​

  • A cross‑bucket explanation
  • A short recommendation
  • Localised high‑value patterns

5.3 Macro-fragments​

  • Multi‑bucket systemic insights
  • Strategic implications
  • Organisational patterns
  • Policy-level understanding

Each fragment becomes a node in a growing knowledge graph.

6. Metadata Schema for Stored Reasoning (Conceptual)

{
  "id": "uuid",
  "buckets": ["incidents", "programs", "transport"],
  "perspective": "reduce anxiety on Wednesday outings",
  "granularity": "meso",
  "created_at": "timestamp",
  "tools_used": ["cluster_v3", "stats_delta", "run_bucket_reasoning"],
  "llm_models_used": ["gpt‑5.1", "local_reasoner_1"],
  "input_window": "last_6_months",
  "output_tokens": 2450,
  "confidence": 0.78,
  "summary_micro": "...",
  "summary_normal": "...",
  "summary_large": "...",
  "insights": [...],
  "cross_bucket_themes": [...],
  "vector_embedding": "512-dim-array"
}

Stored in:

  • Postgres JSONB
  • Vector DB
  • And possibly cached as flat files for fast loading

7. Propagation to Other Brainframe Modules

After saving each new reasoning result, the Propagation Engine checks:

  • Does this insight affect other buckets?
  • Should this trigger a re‑reasoning of other domains?
  • Should new embeddings be generated?
  • Should this adjust risk trends?
  • Should policy or scheduling be rebalanced?

This allows Brainframe to:

Grow in width (more perspectives) and depth (more refined internal knowledge).

8. Dynamic Knowledge Growth

Every interaction can improve the system:

  • New perspectives become part of the brain
  • New correlations are saved
  • New hypotheses are created
  • Old insights fade if they become irrelevant
  • Vector recall surfaces historical connections
  • Consciences cross‑train each other
  • The LLM begins to answer faster because more “pre‑thought” exists

This creates a living organisational intelligence.

9. Diagram

flowchart TD

UQ([User Question])

subgraph FinalStepLLM[Final LLM Call]
    Decide[Evaluate Pre-Reasoned Knowledge]
    NeedMore{Precomputation<br>Insufficient?}
    UsePre[Use Pre-Reasoned Data]
    CallTool[Call run_bucket_reasoning]
end

subgraph Brainframe
    subgraph Buckets[Context Buckets]
        B1[Incidents]
        B2[Programs]
        B3[Transport]
        B4[Scheduling]
    end

    subgraph Reasoner[Reasoning Pipelines]
        FetchRaw[Fetch Raw Data]
        MathOps[Maths & Filters]
        SubLLMs[Secondary LLM Reasoning]
        Summaries[Generate New Summaries]
    end

    Store[Store & Index Outputs]
    Prop[Propagation to Other Buckets]
end

UQ --> Decide
Decide --> NeedMore

NeedMore -->|No| UsePre --> FinalOut[Final Answer]
NeedMore -->|Yes| CallTool --> FetchRaw
FetchRaw --> MathOps
MathOps --> SubLLMs
SubLLMs --> Summaries
Summaries --> Store
Store --> Prop
Prop --> CallTool
Summaries --> FinalStepLLM
FinalStepLLM --> FinalOut

10. Final Definition

Cumulative Fragmented Consciences
= Fragmented minds that think continuously and learn from every question.
They produce structured, storable, recallable reasoning that becomes richer over time.

This is how Brainframe becomes:

  • a living knowledge system
  • a dynamic organisational memory
  • a self-growing intelligence
  • a proactive advisor
  • and a continuously evolving strategic engine