┌─────────────────────┐
                        │    brain-server      │
                        │   (axum, 60+ API)    │
                        └──────────┬──────────┘
                                   │
          ┌────────────────────────┼────────────────────────┐
          │                        │                        │
┌─────────┴─────────┐  ┌─────────┴─────────┐  ┌──────────┴─────────┐
│  brain-cognition   │  │  brain-inference   │  │   brain-spiking    │
│                    │  │                    │  │                    │
│ Working memory     │  │ DINOv2  (384d)     │  │ 10 brain regions   │
│ Hopfield memory    │  │ CLIP    (512d)     │  │ 2M ALIF neurons    │
│ Knowledge graph    │  │ Whisper (512d)     │  │ 2B CSR synapses    │
│ Personal memory    │  │ MiniLM  (384d)     │  │ Three-factor STDP  │
│ Companion/emotion  │  │ World model        │  │ 4 neuromodulators  │
│ Autonomy loop      │  │ Mel spectrogram    │  │ Cell assemblies    │
│ Sleep consolidation│  │ VAD, faces         │  │ Triple extraction  │
└────────────────────┘  └────────────────────┘  │ Chain recall       │
                                                 │ Sleep/pruning      │
          ┌─────────────────────┐                └────────────────────┘
          │    brain-core       │
          │ Hebbian networks    │
          │ Sparse projections  │
          └─────────────────────┘

YouTube video
  → yt-dlp auto-subtitles
  → Sentence chunking (~200 chars, filler filtering)
  → LLM-powered triple extraction (Ollama, batched):
      "From these sentences about X, extract subject|verb|object triples"
      e.g., TurboQuant|compresses|KV cache
  → Batch learning: all triples encoded in one tick (~0ms)
  → Concept association matrix: S→R, R→O, S→O edges strengthened
  → No text stored — association weights ARE the memory

Query → Fuzzy concept matching → Find matching cell assemblies
  → BFS through concept association graph (up to 6 hops)
  → Follow strongest weighted edges, filter noise concepts
  → Chain of associated concepts returned with strength scores
  → Injected into LLM system prompt as learned knowledge

Q: "tell me about TurboQuant"
A: "I don't have specific learned knowledge about TurboQuant..."

Input: https://www.youtube.com/watch?v=7YVrb3-ABYE

→ LLM identifies topic from transcript: "turbo quant"
→ 53 triples extracted (turbo quant|compresses|kv cache, turbo quant|uses|random rotations, ...)
→ Triples stored in spiking brain (synaptic imprinting) + knowledge graph (SQLite)
→ Concepts seeded: turbo quant, lora, kv cache, random rotations, memory usage, ...

Q: "tell me about TurboQuant"

Grounding (learned concepts matched):
  - turboquant (0.58)
  - random rotations (0.48)
  - chopping off end of numbers (0.45)

KG edges retrieved:
  - turbo quant → compresses → kv cache
  - turbo quant → uses → random rotations
  - turbo quant → affects → memory usage

A: "TurboQuant is a system that impacts large semiconductor companies'
    stock prices by efficiently compressing kv cache and reducing memory
    usage through the use of random rotations. It's particularly relevant
    in the context of large language models where it helps optimize
    performance, though its implementation has also been subject to
    controversy."

Q: "what is LoRA?"
A: "I don't have learned knowledge specifically about LoRA..."
   (learned concepts: [])

→ Topic: "lora" (extracted from transcript by LLM)
→ 60 triples extracted
→ Concepts: lora, low-rank adaptation, base model weights, parameter numbers, ...

Q: "what is LoRA?"

Grounding:
  - lora (0.60)
  - LoRA: Low-Rank Adaptation of Large Language Models (0.40)
  - low-rank adaptation (0.39)
  - low-rank decomposition (0.38)
  - read-only to base model weights (0.36)

A: "LoRA stands for Low-Rank Adaptation, a method used in machine
    learning to efficiently adapt large models to new tasks without
    the need for full retraining. It achieves this by updating only a
    small subset of parameters using low-rank matrices, which helps in
    reducing both computational expense and storage requirements."

[explicit] kv cache (strength: 200)           — from BFS
[explicit] short-term memory of models (160)   — from BFS
[explicit] formal mathematical proof (160)     — from BFS
[emergent] sparse (strength: 70)               — discovered by spiking propagation
[emergent] word into vector (strength: 70)     — discovered by spiking propagation
[emergent] similar words close to each other (70) — discovered by spiking propagation

rust/
  crates/
    brain-spiking/     # Spiking neural network engine
      src/
        neuron.rs       # ALIF neurons, SoA layout
        synapse.rs      # COO builder → CSR storage, synaptic scaling
        region.rs       # Brain region (neurons + synapses + STDP)
        network.rs      # Multi-region orchestrator
        concepts.rs     # Cell assemblies, triple extraction
        knowledge.rs    # Concept association matrix, BFS chain recall
        plasticity.rs   # Three-factor STDP, TACOS dual-weight
        neuromodulation.rs  # DA, ACh, NE, 5-HT
        sleep.rs        # NREM replay + REM noise + structural pruning
        spike_encoder.rs    # Latency coding (embedding → spikes)
        spike_decoder.rs    # Rate decoding (spikes → embedding)
    brain-server/      # HTTP server (axum)
    brain-cognition/   # Cognitive systems
    brain-inference/   # TorchScript model loading
    brain-core/        # Hebbian networks
    brain-db/          # SQLite persistence
    brain-experiment/  # Self-improving mutation loop
scripts/               # Python training, data download, model export
templates/             # Web UI (8 HTML pages)
docs/                  # API reference, engineering lessons