Brain Configuration

The Brain is Synapses’ in-process LLM enrichment system. It uses local Ollama models to add semantic understanding to the code graph — auto-summarization, rule explanations, architectural analysis, and multi-agent coordination.

{
  "brain": {
    "enabled": true,
    "intelligence_mode": "optimal",
    "ollama_url": "http://localhost:11434",
    "model": "qwen3.5:2b",
    "ingest": true,
    "enrich": true
  }
}

Fields

Field	Type	Default	Description
`enabled`	`bool`	`false`	Activate LLM enrichment
`intelligence_mode`	`string`	auto	`"optimal"` (8GB), `"standard"` (16GB), `"full"` (32GB+)
`ollama_url`	`string`	`localhost:11434`	Ollama API endpoint
`model`	`string`	`qwen3.5:2b`	Primary model for all tiers
`fast_model`	`string`	`qwen3.5:2b`	Bulk ingestion model
`model_ingest`	`string`	—	Override model for Tier 0 (Reflex)
`model_guardian`	`string`	—	Override model for Tier 1 (Sensory)
`model_enrich`	`string`	—	Override model for Tier 2 (Specialist)
`model_orchestrate`	`string`	—	Override model for Tier 3 (Architect)
`model_archivist`	`string`	—	Override model for archival tasks
`db_path`	`string`	`~/.synapses/brain.sqlite`	Brain database location
`ingest`	`bool`	—	Auto-summarize entities on file save
`enrich`	`bool`	—	LLM enrichment on `get_context` calls
`context_builder`	`bool\|null`	`null`	Context packet building. `null` inherits from `enabled`

Intelligence Modes

Mode	RAM Required	Model	Notes
optimal	8 GB	qwen3.5:2b (~1.5 GB)	Shares identity across tiers. Best for laptops.
standard	16 GB	qwen3.5:4b (~2.7 GB)	Separate guardian identity. Better analysis.
full	32 GB+	qwen3.5:4b	Models pinned in RAM. Fastest responses.

Brain Tiers

The brain operates as a 4-tier pipeline:

Tier 0 — Reflex (Ingestor + Pruner)

Auto-summarizes code entities on file save
Generates 1-sentence summary + 1-3 domain tags per entity
~500ms latency per entity
Strips boilerplate from web content

Tier 1 — Sensory (Guardian)

Generates plain-English explanations for architectural rule violations
Provides actionable fix suggestions
~800ms per violation (cached after first hit)
Circuit breaker falls back if Ollama is slow

Tier 2 — Specialist (Enricher + Context Builder)

Two-pass design: deterministic (~5ms) + optional LLM (~2-3s)
Deterministic pass: SDLC phase detection, complexity scoring
LLM pass: 2-sentence architectural insight + concerns
Builds structured Context Packets with 7 sections

Tier 3 — Architect (Orchestrator)

Multi-agent conflict resolution
Suggests non-conflicting work scope for parallel agents
~1-2s per coordination request

Prerequisites

Brain requires Ollama running locally:

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Pull the default model
ollama pull qwen3.5:2b

# Verify
ollama list

Disabling Brain

If you don’t want LLM enrichment (e.g., air-gapped environment):

{
  "brain": {
    "enabled": false
  }
}

All brain features degrade gracefully — deterministic analysis still works, LLM insights are simply omitted.