96 Memories, 133 Reinforcements: Living Institutional Knowledge Inside OPS

Most AI tools forget everything between sessions. Ours has 96 memories—and knows when two of them contradict each other.

What OPS actually does

We've been building OPS, an AI system that learns from human corrections. Not fine-tuning. Not prompt engineering.

Every time someone approves, edits, or overrides an AI recommendation—whether it's an invoice review, a task assignment, or a drafted communication—the system extracts what changed and why, then stores it as a memory with a confidence score.

What's accumulated after a few weeks of dogfooding

• 96 memories across 7 categories: content patterns, editorial preferences, lead insights, strategy, performance metrics, failure patterns, and human corrections.
• 133 total reinforcements so far.
• Each memory reinforced or decayed based on real usage—not a one-shot capture.

The failure patterns are the most interesting part

One of ours: never open a message by praising someone's past work without first establishing why you're reaching out.

That's not a style preference. It's a judgment call our team learned through trial and error. Now the system knows it too.

What happens when memories conflict

Some of them do. Real organizational knowledge is messy—different contexts demand different approaches.

Most systems silently pick one and move on. OPS flags the conflict instead. Surfacing the contradiction is more useful than pretending it doesn't exist.

This is what institutional knowledge looks like when you make it explicit

Not a wiki nobody updates. Not an onboarding doc from 2022.

Living memory that strengthens every time someone makes a correction—and that every new team member inherits on day one.

What's the equivalent of "OPS memory #47" inside your organization—the judgment call that everyone learned the hard way but never wrote down?