MemArchitect introduces a governance layer that decouples memory lifecycle management from LLM weights to prevent 'zombie memories.'
arXiv · March 20, 2026 · 2603.18330
The Takeaway
It provides a framework for resolving contradictions and enforcing privacy in persistent LLM agents, which standard RAG frameworks treat as passive storage. This is essential for building reliable, long-running autonomous agents in production.
From the abstract
Persistent Large Language Model (LLM) agents expose a critical governance gap in memory management. Standard Retrieval-Augmented Generation (RAG) frameworks treat memory as passive storage, lacking mechanisms to resolve contradictions, enforce privacy, or prevent outdated information ("zombie memories") from contaminating the context window.We introduce MemArchitect, a governance layer that decouples memory lifecycle management from model weights. MemArchitect enforces explicit, rule-based polic