Cognitive Memory for Coding Agents: Reusable Reasoning Across Platforms
Friday, April 17, 2026 · 2:35 PM – 3:05 PM · Auditorium 5
AI coding agents solve complex problems, build mental models of codebases, and develop sophisticated debugging strategies. Then the session ends, and all that reasoning vanishes. The next agent rediscovers the same patterns, hits the same pitfalls, and re-derives the same solutions. This context amnesia is one of the biggest practical bottlenecks in AI-assisted development, and it matters now because multi-agent, multi-platform workflows are becoming the norm rather than the exception.
This talk is for developers, ML engineers, and technical leads who use AI coding agents daily and have felt the frustration of repeating context across sessions and tools. Familiarity with at least one AI coding agent and a general understanding of embeddings and vector search is helpful, but no specialized knowledge of memory architectures is required. We build intuition from practical scenarios before introducing the technical framework.
We present a system that extracts reasoning patterns, not just facts, from completed coding sessions across platforms and makes them available to any future agent on the same codebase. The central ideas include: an extraction pipeline that parses heterogeneous session formats and constructs directed acyclic graphs of each agent’s reasoning trajectory, identifying pivotal moments where direction changed, and insights emerged, while filtering roughly 90% of raw noise. A semantic deduplication and compression layer that clusters related memories by embedding similarity, uses LLM-powered merge decisions to eliminate redundancy, and consolidates accumulated knowledge into structured cognitive profiles containing architectural models, ranked pitfalls, and proven diagnostic strategies. A three-tier temporal architecture with hot retrieval during sessions, warm heuristic extraction between sessions in under a second, and cold batch processing for thorough consolidation, solving the real problem of developers bouncing between tools throughout the day. Finally, a cross-platform delivery via the Model Context Protocol.
Attendees will leave with: a concrete architecture for persistent cross platform agent memory they can implement using open source tools with no modifications to any coding agent; the distinction between content memory and cognitive memory with real examples showing why storing how an agent reasoned is more valuable than storing what it concluded; and the three tier temporal pattern applicable to any agent memory system.
About the Speakers
Sazan Khalid
Masters in Data Science at Georgetown
Data Scientist and Masters Student at Georgetown University
Amit Arora
Principal Solutions Architect @AWS, Adjunct Professor Georgetown University
Amit Arora is an AI and ML Specialist Architect at Amazon Web Services, helping enterprise customers use cloud-based machine learning services to rapidly scale their innovations. He is also an adjunct lecturer in the MS data science and analytics program at Georgetown University in Washington D.C.