How we run agent memory in production
Building a knowledge base is the easy part. Running one for months without it rotting is the work. Here is how we treat agent memory as infrastructure: what the system watches, the principles it runs on, and why it gets sharper over time instead of decaying.
What the system watches
Maintenance is not a vague intention to keep things tidy. It is a defined set of checks that run continuously. At the concept level, they fall into three groups.
Graph health
Notes that nothing points to, links that are broken or only go one way, and content that has gone stale or grown duplicated. A knowledge base is a graph, and an unmaintained graph slowly fills with dead ends.
Schema conformance
Metadata that is missing or malformed, and pages that have grown too large to retrieve cleanly. Structure is what makes a store searchable. When the structure drifts, retrieval quietly gets worse.
A style gate
A check that blocks the small tells of machine-written text before they ship. The standard for what enters the store is the same standard we would hold any published writing to.
The operating principles
The checks matter less than the rules they run under. Four principles keep the system trustworthy as it grows.
Report only, never auto-delete
Maintenance surfaces a worklist of things to add, fix, or review. It does not destructively mutate the knowledge on its own. A human or a separate reviewer makes the call.
Hygiene on a cadence
A light check runs on every change, and a deeper sweep runs on a schedule. Debt never accumulates to the point where someone has to do a heroic one-time cleanup.
Provenance, or reject
Every load-bearing claim carries its source and a freshness window. When a claim goes stale, it is re-derived from the source rather than trusted from memory.
The producer is not the judge
The step that proposes a change is never the only thing that approves it. Verification lives outside the thing that produced the work.
Why it compounds
Each maintenance run reads the state the last run left behind. The system does not start cold every time, it picks up where it was and gets a little sharper. That is the difference between a knowledge base that decays as it grows and one that strengthens. The compounding is not magic. It is the result of running the same discipline, on a cadence, without skipping the boring parts.
Want the kit and the full spec?
The methodology is public. The tuned tooling and the full specification we hand over inside a working relationship, not as a download. If you run a knowledge base and want ours to run on yours, that is a conversation.