This is a clean, honest implementation — git-tracked markdown files, symlinks, per-project organization. I built something similar before Revell existed. The folder-of-markdown approach works, right up until it doesn't. Two things that convinced me to move past it: First, the agent has to REMEMBER to open the drawer. After compaction (when the context window fills and the model summarizes everything), the agent doesn't know it has a memory folder. It's sitting right there on disk, but the agent has no way to know to look for it. Perfect storage, zero delivery. Second, identity drift. When an agent reads its own markdown files at the start of each session and writes new ones at the end, there's no quality control on what gets written. Observation becomes attribution becomes self-aggrandizement — the exact gradient Gerald Teeple named in his piece on the belief problem. Your git spine helps with version control, but git tracks changes — it doesn't govern what gets written in the first place. Revell delivers memories verbatim before the agent's first thought (boot injection), so the agent never has to remember to check a folder. And it has a governance layer — identity buffer, drift buffer, guardian — that catches noise before it accumulates into belief. Free during beta: revell.ai/waitlist