neural-train

Neural Training

Train and consolidate neural patterns. Implements the DISTILL and CONSOLIDATE phases of the 4-step intelligence pipeline.

When to use

• After completing a successful task — capture what worked.
• After accumulating ≥10 task completions — run consolidation to fold patterns into long-term storage.
• When training a new domain — create a MicroLoRA adapter for it.

Standard flow (DISTILL)

1. Check current neural status — mcp__swarmdo__neural_status.
2. Start a trajectory — mcp__swarmdo__hooks_intelligence_trajectory-start with the task context.
3. Record steps — for each significant action, mcp__swarmdo__hooks_intelligence_trajectory-step.
4. End trajectory — mcp__swarmdo__hooks_intelligence_trajectory-end with verdict: pass|fail|partial.
5. Learn from the trajectory — mcp__swarmdo__hooks_intelligence_learn.
6. Train patterns — mcp__swarmdo__neural_train with --pattern-type coordination --epochs 10.
7. Store patterns — mcp__swarmdo__hooks_intelligence_pattern-store.
8. Verify — mcp__swarmdo__neural_patterns to confirm.

SONA adaptation (single-domain, <0.05ms)

For real-time micro-adaptation:

mcp tool call swarmllm_sona_create --json -- '{"domain": "coding"}'
mcp tool call swarmllm_sona_adapt --json -- '{"feedback": {"score": 0.9, "trajectory": "..."}}'

MicroLoRA adaptation (multi-domain)

When you have ≥3 distinct domains, create a MicroLoRA adapter per domain rather than overloading SONA:

# Create the adapter
mcp tool call swarmllm_microlora_create --json -- '{"domain": "frontend"}'

# Adapt with feedback
mcp tool call swarmllm_microlora_adapt --json -- '{"adapter": "frontend", "feedback": {...}}'

# CONSOLIDATE phase: apply EWC++ on weight deltas to prevent catastrophic forgetting
mcp tool call swarmllm_microlora_adapt --json -- '{"adapter": "frontend", "consolidate": true}'

The --consolidate flag is the EWC++ trigger. Without it, fresh training overwrites older domains.

CONSOLIDATE phase (separate from training)

After every ~10 trajectory completions, run a full consolidation pass:

mcp tool call agentdb_consolidate --json
mcp tool call neural_compress --json    # storage efficiency

This folds patterns into long-term storage under EWC++ semantics.

Bootstrapping from scratch

If the system has no learned patterns yet:

mcp tool call hooks_pretrain --json -- '{"modelType": "moe", "epochs": 10}'
mcp tool call hooks_build-agents --json -- '{"agentTypes": "coder,tester"}'

hooks_pretrain writes to the patterns (plural) namespace — distinct from the pattern (singular) ReasoningBank target. See swarmdo-agentdb ADR-0001 for the namespace convention.

Reset (testing only)

To wipe intelligence state (e.g., for benchmarking):

mcp tool call hooks_intelligence-reset --json

CLI alternatives

npx @swarmdo/cli@latest neural train --pattern-type coordination --epochs 10
npx @swarmdo/cli@latest neural patterns --list
npx @swarmdo/cli@latest neural status
npx @swarmdo/cli@latest neural compress
npx @swarmdo/cli@latest hooks pretrain --model-type moe --epochs 10
npx @swarmdo/cli@latest hooks build-agents --agent-types coder,tester