nima-core_skill

Overview

This skill implements Neural Integrated Memory Architecture (NIMA) — a production-ready, graph-based memory system for AI agents with semantic search, lazy recall, and dynamic affect. It supports SQLite for development and LadybugDB for high-performance vector search and graph traversal. The system is designed for thread safety, security hardening, and scalable recall injection into agents.

How this skill works

NIMA captures three-layer memory on every turn (input, contemplation, output), stores nodes and edges in a local database, and computes embeddings for semantic indexing. A recall hook injects relevant memories lazily before agent start, deduplicating injected context. A dynamic affect engine processes text in real time using a Panksepp-inspired 7-affect model to modulate response style.

When to use it

• Add persistent, context-aware memory to conversational agents
• Improve retrieval relevance with vector search and graph queries
• Inject concise, deduplicated memories before agent execution
• Add emotion-aware response modulation for more natural interactions
• Migrate from a simple transcript log to structured memory graph

Best practices

• Use LadybugDB in production to enable native HNSW vector search and Cypher graph traversal
• Limit lazy recall results to the top-N to keep agent context compact and relevant
• Configure NIMA_EMBEDDER to a local provider when you must avoid external network calls
• Store agent transcripts under a dedicated NIMA_DATA_DIR to simplify backups and audits
• Set sensible API timeouts and rotate embedding keys regularly

Example use cases

• Customer support agent that recalls prior tickets and customer sentiment to personalize replies
• Research assistant that retrieves past notes and relationships via graph traversal
• Personal assistant that remembers user preferences and modulates tone based on affect state
• QA system that searches archived agent sessions across multiple versions for troubleshooting
• Hybrid local/cloud deployments that use local embeddings for privacy and Voyage/OpenAI for quality when permitted