Stop hardcoding memory for your AI agents—because now they can design it themselves. In this video I break down MemEvolve, an open-source approach where the agent doesn’t just “store more memories,” but actually evolves the full memory architecture over time. The key problem with most agent memory today is staticity: we pick a fixed design (RAG, keyword search, hybrid search, tool calls, graphs, summarization, etc.) and hope it fits the use case. But the real question is: how do you know which memory architecture is best for your specific agent and workload? MemEvolve tackles this by defining memory as four components—encode, store, retrieve, and manage—and then running an evolutionary loop to test competing candidates on your task examples (think 50–150 queries). It keeps the best-performing candidate based on metrics like accuracy, cost, and latency, generates a diagnosis report from the logs, and then uses an LLM to implement a better architecture for the next iteration. The result is a specialized memory system tuned to your use case, and you can re-run the “freezer stage” whenever your production traffic or tasks change. The paper reports big gains (often ~20% on average and up to ~70% in some settings) versus prior state-of-the-art memory designs, which is why I think this is a really important shift for agentic systems.