agent-safla-neural

Use the SAFLA four-tier memory model combining vector, episodic, semantic, and working memory layers. Store learning patterns via the mcp__claude-flow__memory_usage tool with namespaced keys and TTL settings so context survives across sessions.

Call mcp__claude-flow__neural_train with a pattern_type such as coordination, training data describing the architecture and memory tiers, and an epoch count. The skill example uses 50 epochs with feedback loops and persistence enabled.

It is an architecture where agents store interaction outcomes, extract learned patterns with confidence scores, and feed them back into future decisions. This creates self-improving cycles that adapt strategies based on performance metrics.

Yes, the architecture supports swarm memory sharing, coordinating distributed memory across agent swarms. Stored patterns use namespaces and keys, allowing multiple agents to read and write shared learning state.

Feedback loops depend on the quality of stored outcomes and confidence scoring, so poor initial metrics can reinforce bad patterns. The skill addresses this with safety constraint frameworks, but memory storage also requires TTL management to avoid stale context.