Dev Tools · 1h ago
Dev builds benchmark to test markdown as agent memory, hits 96% of hand-built ceiling
The IWE knowledge tool team created a benchmark to test whether plain markdown files can serve as AI agent memory. After iterating on search, editing, and curation, a $4.50 model achieved 96% accuracy versus a hand-built ceiling. Grep on raw transcripts still beats all methods on accuracy alone.
Meridian48 take
The story is a deep dive into iterative engineering, but the headline result—96% with a cheap model—is less surprising given the benchmark's narrow scope and the fact that grep remains unbeaten.
agent-memorymarkdown-knowledge-graph