We build agents to solve problems specific to African markets — iTenda is the clearest example, turning tender bidding from a manual slog into something an SME can actually keep up with. Underneath that work is a smaller, recurring problem we kept running into ourselves: every session with an AI agent produces hard-won lessons — corrections, workarounds, methods that took a few tries to land — and most of it evaporates the moment the session ends.
Memory systems solve part of that: they remember facts about you. They don't solve the other part — how to reliably do a specific kind of task. That's what agent skills are for, and it's the gap most memory tooling leaves untouched.
So we built Upskill — a small, open Agent Skill that closes the loop. At the end of a working session, it looks back for whatever actually recurred — not one-off details, but corrections and methods likely to come up again — checks whether an existing skill already covers that ground, and updates it, or creates a new one, rather than letting the lesson die with the conversation. It's domain-agnostic: useful for a developer, a researcher, a lawyer, anyone doing repeatable work with an agent's help.
It's a small release, but it's exactly the kind of thing we mean when we say our approach is grounded in research, not assumptions: we noticed the same friction recurring across our own work, tested it against a real scenario before trusting it, and found — and fixed — a real gap in the process. We're sharing it because the problem clearly isn't ours alone.
Upskill is free and open source (MIT), available now.
View on GitHub