Descriptions:
At AI Dev 26 in San Francisco, Carter Rabasa — Head of Developer Relations at Box — makes the case that file systems are the most practical and underappreciated primitive for giving AI agents persistent, long-term memory. With 90% of enterprise data living in unstructured files (PDFs, Word docs, spreadsheets), Rabasa argues that cloud file systems like Box are already purpose-built for the retrieval, update, and inspectability requirements that agentic memory demands — without needing bespoke vector databases or custom APIs.
The talk centers on a live demo using the OpenAI Agents SDK’s remote file system mount feature, which allows sandbox agents to treat a Box folder as if it were a local file system. Rabasa walks through a team-management agent built in fewer than 20 lines of Python — backed by two Markdown files in Box — that can read, update, and reason over task assignments across sessions. The agent prompt, skill definition, and file structure are all shown explicitly, making the architecture directly reproducible.
Rabasa outlines four requirements any long-term memory solution must meet: cross-session fact retention, selective retrieval, updateability, and human inspectability. He argues that file systems satisfy all four naturally, and that LLMs already understand file system semantics deeply — making the integration unusually low-friction compared to embedding pipelines or graph databases. The talk is aimed at developers building production agents who need durable context without significant infrastructure overhead.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Hands On Build
