Anthropic is developing Memory Files, an upcoming update to Claude’s memory system where notes are distributed across multiple structured documents organized by topic, project, or context, Testing Catalog reported on May 22 based on UI traces.
The architectural shift is from a single growing conversational memory to a sharded structure where the model can retrieve a targeted slice of context for each task. The bottleneck on long-running agent memory has always been retrieval signal-to-noise: as memory accumulates, finding the relevant piece for the current task becomes harder, and the model’s attention budget is consumed by retrieval rather than reasoning. Sharding context by topic gives the model a more targeted lookup, which trades the simplicity of one memory blob for the precision of structured retrieval.
The pattern echoes Cursor’s “rules” system, where developers maintain separate context files for different concerns, and Claude Code’s recent shift from Markdown to HTML for its default context format (richer structure, better disambiguation). The broader trend is structured-context-over-flat-context across the agentic tooling stack.
For teams running Claude-backed agents in production, Memory Files will reshape how persistent context is managed. The migration path from a flat-memory deployment to a sharded one is non-trivial, and the structure of the topic-or-project boundaries will need to map onto the actual semantic categories your agent operates over. The 6-month planning horizon should assume Memory Files lands during that window.
Reported by Testing Catalog on 2026-05-22.
