MUMBAI, India, July 13 -- Intellectual Property India has published a patent application (202641043277 A) filed by Sundaraparipurnan Narayanan; and Marie Potel on April 04, 2026, for “a Dynamic Signal-Driven Memory Architecture For Large Language Model Agents”.
Inventors include Sundaraparipurnan Narayanan; Marie Potel; and Sandeep Vishwakarma.
The application for the patent was published on July 10, 2026, under issue no. 28/2026.
Abstract: ABSTRACT A Dynamic Signal-Driven Memory Architecture for Large Language Model Agents The present invention relates to artificial intelligence systems and, more particularly, to a dynamic signal-driven memory architecture for large language model agents. The architecture receives memory entries from reasoning or alignment systems and classifies each entry into one or more of six adaptive categories comprising Short-term, Long-term, Session/Episodic, User-centric, Enterprise/Org/Use Case Context, and Temp/Transient memory. Four real-time metadata signals comprising Frequency, Relevance, Criticality, and Learning Rate are computed for each memory entry and stored in a structured repository. A memory signal monitor tracks signal evolution and cooperates with a trigger and scheduling module to invoke a realignment LLM agent periodically or upon threshold breach. The realignment LLM agent promotes, demotes, copies, or archives memories according to contextual utility. During inference, a retrieval and purge module computes weighted top-k rankings so that only the most useful memories are supplied to the agent. Obsolete memories are archived or purged with unique link retention for traceability and audit reconstruction. The invention thereby reduces storage bloat, decreases retrieval noise, improves context-aware performance, and supports regulated and enterprise-grade deployments.
Disclaimer: Curated by HT Syndication.
