Constrained Autonomy: AI Models, Memory, and Retrieval at the Edge

An evening of three technical talks on the engineering realities of running AI on constrained physical systems and edge devices. The program splits the on-device stack into three layers, models, memory, and retrieval, and works through each in turn: what it costs to run an LLM locally, how a long-running agent keeps track of its own state, and how vector search can happen inside the hardware rather than in a datacenter. Paul Willot (Senior MLE, Liquid AI) opens with the economics of inference, looking at how memory bandwidth, model size, hardware utilization, and concurrency set the cost and speed of a token, and at how quantization, sparsity, hybrid architectures, caching, and speculative decoding shift that trade-off. Stefania Druga (Staff Research Scientist, Sakana AI RSI Lab) follows with agentic memory, drawing on experiments with research agents that run for hundreds of turns to compare memory-off, deployed recall, gating, and active sub-goal ranking, and to argue that memory is a policy rather than storage. Ewa Szyszka (DevRel Engineer, Qdrant) closes with Qdrant Edge, an embedded vector search library with an install footprint around 11 MB, demonstrated in a live robot demo where hybrid search over dense vision vectors and BM25 captions returns in under a millisecond. Doors open at 18:00, talks run 18:30 to 20:00, and networking follows until 21:00. The material is aimed at engineers working in robotics, applied ML, and embedded systems who need agentic behavior that stays reliable without a cloud connection.