A memory and governance layer that makes the whole agent fleet smarter
outcome sessions start already knowing the sharp edges
Problem
Two failure modes eat agent productivity: the AI forgetting every environment quirk between sessions, and skill sprawl where dozens of tools fight over the same job while the human drowns in approval prompts.
What I built
A persistent atomic memory: over a hundred single-fact files with typed frontmatter, cross-linked, with a one-line index loaded every session, and written hygiene rules (trust the ticket system over memory when they disagree, prune shipped work, dedup before writing).
On top of it, a governance layer in the workspace instructions: which skills own workflows versus which are technique-level tools, explicit do-not-invoke lists where a project skill supersedes a generic one, and interview etiquette (batched questions with recommendations, one approval per batch, act-then-report for reversible actions).
Outcome
Sessions start already knowing the environment’s sharp edges, and human attention is spent on genuine forks rather than routine confirmations. The approach transfers directly to any team adopting agent tooling at scale.