Spiking Recall Resurrection
Fire BFS-discovered concepts into the 2B-synapse spiking network
30-step propagation through association cortex with raised synaptic clamp (1.5)
Neuromodulator-driven recall: focused mode (high ACh) for single-topic, broad mode (high NE) for cross-domain
Emergent associations discovered through lateral neural pathways
Example: querying "TurboQuant" activates "sparse", "word into vector" from different topics
Confidence tags: [confirmed] (BFS+spiking), [explicit] (BFS only), [emergent] (spiking only)
Synaptic Knowledge Imprinting
When triples are learned, strengthen actual CSR weights between concept assembly neurons
803 synapses imprinted per 13-triple learning batch (delta 0.8/0.8/0.4)
Persisted triples re-imprinted on startup
Imprinted weights cap at 1.0 (2x the normal 0.5 background)
Persistent Cumulative Knowledge
Triples persist to data/triples.log (pipe-delimited, append-only)
Topic registry in data/topics.json (deduplication, provenance)
Knowledge survives server restarts: 24 topics, 816 concepts, 1195 associations
Weight cap raised to 2.0 for multi-source reinforcement
Bidirectional BFS: query matches 2+ concept clusters, BFS from both sides, find bridge nodes
Bridge concept detection: "kv cache" shared between TurboQuant and FlashAttention
Cross-domain answers: LLM connects knowledge from different YouTube videos
Topic provenance tracking on ConceptRegistry
POST /api/brain/learn/batch — learn multiple videos {"videos": [{url, topic}, ...]}GET /api/brain/knowledge/stats — topics, concepts, associations, bridges, top connected
LLM-Powered Triple Extraction
Replaced rule-based SVO parser with LLM-powered extraction (Ollama)
Batched sentences (10 per call) for efficient extraction
Example: "TurboQuant compresses the KV cache" now extracts TurboQuant|compresses|KV cache
Noise filtering: rejects prompt echoes, long rambling objects, meta-text
Rule-based fallback when Ollama is unavailable
Triple queue now drains all pending triples in a single tick
12 triples learned in 0.000s (previously 48s at 1 per 2s tick)
Filtered relation verbs ("is", "are", "relates-to") from recall output
Only substantive concept names returned as associations
System prompt now instructs LLM to use brain knowledge as factual learned data
LLM answers are grounded in brain associations, not hedged guesses
Triple Extraction Quality (Rule-Based Fallback)
Sentence-boundary splitting for multi-sentence chunks
Expanded noise filters: filler subjects, commas, single junk words
Topic-anchored extraction: catches key phrases near topic even without SVO
Stop word list expanded (50+ words)
Direct Concept Association Matrix
Replaced 500M-synapse STDP simulation with HashMap-based associations
learn_triple: 3 hash map updates (S→R, R→O, S→O edges)
recall_chain: BFS through association graph (instant)
Learning went from 90s/triple to 0ms/triple
Optional CUDA acceleration via tch/libtorch
COO format on GPU, scatter_add spike delivery
Feature-gated: --features gpu
Performance Optimizations
CSR prefetching (x86_64 _mm_prefetch)
Sorted spike delivery for cache locality
target-cpu=native for SIMD auto-vectorizationThread-local reusable buffers for synaptic delivery
10 brain regions, 2M ALIF neurons, 2B CSR synapses
Three-factor STDP with eligibility traces
TACOS dual-weight synapses for continual learning
4 neuromodulators (dopamine, acetylcholine, norepinephrine, serotonin)
Cell assemblies (~100 neurons per concept)
Foundation model encoders: DINOv2, CLIP, Whisper, MiniLM
YouTube video learning pipeline
60+ API endpoints via axum
Sleep consolidation (NREM replay + REM noise + structural pruning)
PolyForm Noncommercial 1.0.0 license