Ignacio Martinez

Ignacio “Nacho” Martinez is an AI developer advocate at Oracle who teaches developers to build machine-learning applications and agents through working code, hands-on workshops, and open-source projects. His teaching spans gaming-based machine learning and persistent agent memory. With Richmond Alake, he co-teaches Oracle and DeepLearning.AI’s Agent Memory: Building Memory-Aware Agents, which shows developers how to retain and update useful information across agent sessions.

Teaching machine learning through games

Earlier in his Oracle teaching work, Martinez used League of Legends to make the machine-learning workflow concrete. Listed as a data scientist advocate, he co-presented three hands-on workshops with Victor Martin; Wojciech Pluta also joined the data-extraction session. The series moved from collecting data to building models with scikit-learn and AutoGluon, then introduced neural networks. A familiar game gave learners a setting in which to understand the steps between raw data and a trained model.

His open-source portfolio carries that practical approach into racing telemetry, computer vision, and deep learning. It includes Formula 1 telemetry integration, YOLOv5 mask-detection tutorials covering labeling, training, and real-time inference, and guides to working with data in PyTorch and TensorFlow. These examples fit his stated interest in teaching AI in ways that help people understand how its pieces work together.

His agent work includes a personal implementation of reasoning capabilities, a PicoClaw adaptation for Oracle AI Database, and educational material on combining the Agent2Agent protocol with LangChain in a multi-agent retrieval system. The projects extend his teaching into the software around language models: reasoning, communication between agents, retrieval, and persistence.

Making agent memory concrete

Released in March 2026, Martinez and Alake’s agent-memory course addresses a practical limitation of stateless agents: a new session can discard useful context from earlier interactions. The course treats long-term memory as structured infrastructure outside the model, giving developers mechanisms for deciding what to preserve, how to retrieve it, and when to revise it.

Its hands-on modules use LangChain, Tavily, and Oracle AI Database to develop several connected capabilities:

• Coordinate memory during execution. Learners design persistent stores for different memory types and implement a manager that handles reading, writing, and retrieval. Oracle AI Database combines vector search for semantic similarity, graph traversal for relationships, and relational queries for structured information within one engine.

• Retrieve tools as procedural knowledge. Semantic tool memory indexes tools in a vector store so an agent can retrieve those relevant to its current task. This gives developers a way to select tools at inference time without putting the entire collection into the model’s context.

• Extract and revise remembered information. LLM-powered pipelines turn interactions into structured facts, consolidate episodic information into semantic memory, and write updates back to the store. The exercises include handling contradictions so memory can change as new information arrives.

• Carry context across sessions. The final agent loads long-term memory at startup, checkpoints intermediate reasoning states during execution, and persists learned context for subsequent sessions. This connects the individual memory operations into a working agent.

Martinez emphasizes implementations developers can inspect and adapt: memory stores, extraction pipelines, and code that handles conflicting information. Across his gaming workshops and agent education, the recurring teaching method is to turn an unfamiliar AI capability into a sequence of concrete steps, then show how those steps fit into a working system.