humanoid-memory

Long-term memory consolidation for AI agents is implemented using V-score consolidation combining coherence, significance, noise, and resonance to decide what to store across conversational sessions. Tensor fusion then merges similar memories while HNSW-backed semantic search provides fast vector retrieval.

V-score consolidation is an evaluation mechanism that combines coherence, significance, noise, and resonance to decide what memories to store. It filters conversational memory by LLM-based C/S/N evaluation to retain important user facts and discard noise.

Automate memory lifecycle management with exponential forgetting by applying the Ebbinghaus forgetting curve to trigger exponential decay. This automatically purges decayed or low-value memories from per-agent persistent storage during sessions without manual intervention.

Humanoid-memory works with HNSW semantic search and vector embeddings to provide fast vector retrieval. It uses tensor fusion to merge similar memories and HNSW-backed semantic search to inject relevant prior interactions into agent prompts.

You can use this memory system for context injection into LLM prompts via hooks that automatically inject relevant prior interactions and recalled memories into requests. This enables conversational agents to track user preferences across sessions.

The best way to track user preferences in conversational agents over time is combining per-agent persistent storage with V-score consolidation and the Ebbinghaus forgetting curve. This retains important user facts while automatically purging decayed memories.