Beyond Context Windows: Building Cognitive Memory for AI Agents

AI agents are rapidly changing how users interact with software, yet most agentic systems today operate with little to no intelligent memory, relying instead on brittle context-window heuristics or short-term state. This limitation fundamentally constraints personalization, reasoning, and long-term adaptation.

To build agents that function as long-term personalized assistants, the memory problem must be addressed head-on.

In this talk, we introduce LinkedIn’s Cognitive Memory Agent (CMA), a horizontal memory platform designed to power stateful, context-aware, and personalized AI agents at scale. We present the architecture and design principles behind CMA, and how it enables agents to accumulate, reason over, and act upon long-term user knowledge. CMA tackles the memory challenge through three core components.

• An ingestion layer that determines how to interpret unstructured inputs, what information to extract, and when and how to store it
• A layered memory system comprising semantic memory (structured knowledge), episodic memory (time-indexed events), working memory (in-session context), and procedural memory (reasoning traces and plans).
• A retrieval orchestration layer that disambiguates user intent, dynamically retrieves relevant memories across layers, and synthesizes responses.

Reinforcement learning continuously optimizes both ingestion and retrieval for quality, efficiency, and adaptability.

Together, these components enable agents to move beyond simple recall of prior interactions. CMA continuously ingests signals from user behavior and environment, infers latent preferences, and applies reasoning to deliver proactive, personalized assistance.

We conclude with insights from deploying CMA in large-scale products such as the LinkedIn Hiring Assistant, and explore how these experiences point toward a new generation of intelligent AI agents.