God of Prompt reposted this
carrying on the takeways from the panel we did with Carlos Calva Cecilia MoSze Tham & Timmy Ghiurau and building on top of them: "The evolution from memory to experience: Over the past six months, as agentic tools like OpenClaw, Claude Code, and Codex have accelerated from experiments into everyday workflows, we've seen an explosion of "memory" solutions for AI agents: markdown files, local databases, preference logs, project instructions, retrieval layers, startups, and frontier-lab features. They all orbit the same problem: agents are still brittle over time. Memory tries to solve the agent's amnesia problem, but amnesia is only the first layer." - Carlos Building proactive agents requires more than a good model and a smart harness. The agent needs to learn from feedback, not just respond to it. It needs to improve across sessions, behave differently tomorrow than yesterday, and eventually anticipate needs without being prompted. Earlier this year we published SmartSearch, which showed retrieval can be fast, accurate, and cheap: 93.5% accuracy on CPU, with no LLM in the retrieval loop. But retrieval is only the first step. The hard problem is what happens after: deciding what matters, what becomes knowledge, what gets forgotten, and what changes behavior. - Timmy In the human brain, memory is intrinsically bound to emotion and survival — emotions serve as our biological filtration system, signaling what to drop and what to compress into durable behavior. By transitioning AI from static retrieval layers to a learned controller that decides what to consolidate or forget, we are essentially building a digital architecture modeled after the mammalian hippocampus. The future of AI doesn't belong to the models that archive the most data, but to the proactive agents that use memory as a simulator. This is the shift from Deep Research to Deep Imagination, where an agent doesn't just recall the past, but actively traverses a labyrinth of future possibilities to adapt to our needs before we even prompt them. - Cecilia The question I'm sitting with (especially as I am building Prompt Copilot): "What are the core eval criteria for a product with a persistent experience layer?" If memory ≠ experience, accuracy on retrieval benchmarks isn't enough. We need a way to measure whether the agent actually behaves differently tomorrow because of what happened today. Whoever ships that eval framework first probably shapes the next two years of this space.