Blog
Comprehension debt finally has a name: the gap between code that exists and code anyone understands. Better specs don't close it. Versioned edges between specs, evidence, and code do.
AI agent memory systems store preferences as flat files with no relationships. When context changes, nothing flags stale memories for review. We noticed this while building our own site — and realized it's the same problem Lattice solves for code.
Your agent passes every lint check, writes clean tests, and ships code that solves a problem nobody asked it to solve. The failure isn't in execution. It's in intent.
Context engineering is the term of the moment. But most approaches focus on runtime — what fills the context window for each LLM call. The missing piece is upstream: where does the knowledge come from, and is it still valid?
Lattice gives you traceability. But traceability assumes you know what you're looking for. QMD adds the missing piece: semantic search over your knowledge graph, running entirely on-device.