The Inference Report

Agents and long-term memory are consolidating around structured knowledge graphs and skill inventories. Deer-Flow and Cognee represent the emerging pattern: autonomous systems need persistent state across sessions, not just prompt-level context windows. DeusData's codebase-memory-mcp takes this further by indexing code as a queryable knowledge graph, reducing token overhead by 99 percent and returning results in sub-millisecond time. The practical win here is obvious, agents that can reason over their codebase without burning context tokens on raw file contents. Anthropic-Cybersecurity-Skills and mattpocock's skills repository show the flip side: frameworks standardizing how agents discover and invoke capabilities. Both repos map skills to formal threat models or engineering domains, making agent behavior auditable and composable rather than buried in natural language prompts.

Video production tooling is fragmenting between AI-first and design-first approaches. OpenMontage positions itself as agentic, offering 52 tools and 500 agent skills to orchestrate production workflows. Palmier-Pro takes the opposite stance, a native macOS editor designed for human-AI collaboration in video editing itself. Penpot's dominance in design tools shows that open-source alternatives to proprietary SaaS gain traction when they preserve the core workflow users expect, not when they try to reinvent it. The broader signal: developers are building infrastructure for AI agents to operate on creative tasks, but the UI layer still matters. Agents need scaffolding, and that scaffolding either integrates into existing tools or replaces them entirely, there's little middle ground.

Jack Ridley