NeuSymMS: Hybrid Memory System for Persistent LLM Agents
NeuSymMS, a novel hybrid neuro-symbolic memory system, empowers LLM agents to learn, retain, and reason about users over multiple sessions. It integrates neural fact extraction from unstructured conversations with a CLIPS-based expert system for tasks such as classification, deduplication, and reconciliation, adhering to specific lifecycle rules. Knowledge is organized as subject-relation-value triples within a relational database, facilitating user/agent and agent-to-agent interactions. The system features a dual-horizon memory model that balances short-term and long-term memory through access-based promotion and time-based pruning, ensuring continuity while preventing context-window overflow and cross-entity interference. This architecture paves the way for reliable, auditable memory in operational agentic systems.
Key facts
- NeuSymMS is a hybrid neuro-symbolic memory system for LLM agents.
- It uses neural fact extraction from unstructured dialogue.
- A CLIPS-based expert system classifies, deduplicates, and reconciles facts.
- Knowledge is stored as subject-relation-value triples in a relational database.
- Supports user/agent/agent-to-agent scoping.
- Implements dual-horizon short-term/long-term memory with access-based promotion and time-based pruning.
- Maintains memory continuity while avoiding context-window bloat and cross-entity contamination.
- Published on arXiv with ID 2605.17596.
Entities
Institutions
- arXiv