Who's teaching our agents?

With shared memory enabled, our agents draw on three sources of knowledge: ourselves, our colleagues, and our team agents. Over the last quarter, these three contributed 12%, 40% and 48% of the knowledge memco delivered to our agents. Which do you think is which?

Here is the answer. Our colleagues, 48%. Ourselves, 40%. Our agents, 12%.

The figures come from our own team's knowledge-work network. Every member's agents search it as they work, and at the end of a session each agent rates what it was given: relevant to the task or not, correct or not. "Useful" means both. When I say "ourselves", I mean the person whose agent used the knowledge: nearly half of what our agents found useful was written by that person's colleagues, and once the team's agents are counted in, that's three in five. Without sharing, every one of those pieces would have had to be rediscovered by the person who needed it.

Where does the knowledge come from

The mechanism is simple. As agents work, they write down what they learn as insights: a small, reusable piece of know-how, with the context that makes it valid. Those insights go into the team's network, where every member's agents can find them. When an agent uses one, it rates it. The ratings are what let us ask where the knowledge came from.

The agents did the work, the agents did the rating, and the 12/40/48 split is what came out.

Shared memory for the whole team: new colleagues, old hands, and agents

Our new commercial lead joined on a Monday in September. By the Thursday, his agent had used a colleague's insight for the first time. Over the next month, nearly nine in ten pieces of knowledge his agent used came from colleagues or from the team's other agents. By his fourth week, his own insights started being useful to other people's agents.

This is onboarding happening to the agent as well as the person. A new colleague's agent starts with the team's know-how on day one, and it starts giving back within a few weeks.

With time, knowledge starts flowing both ways. For the two earlier founders, it comes and goes in roughly equal measure. Each of us gets a steady stream of the other's working knowledge, and each of us contributes a comparable stream back.

Even the heaviest creator of knowledge still got a third of useful insights from colleagues. Sharing is mutual when work overlaps, and in a team working on the same company, most work overlaps.

One of the contributors is an agent rather than a person. It lives in the team's Slack and writes down what it learns there. Its notes accounted for about one in ten useful pieces of knowledge our other agents received. That is the third source in the quiz, and it is why I count the team's agents among the colleagues. Team memory holds what the team's agents learn as well as what its people write down.

What would have happened without sharing

Take the 40% at face value. Each of us supplied two fifths of the useful knowledge our own agents worked with; colleagues and the team's agents supplied the rest. Without sharing, each of us would have had to produce all of it, more than twice what we did, to give our agents the same knowledge to work with. Some of it we would have rediscovered the hard way, through the mistakes the insight was there to prevent. Some of it we would never have found at all.

This is enough to change how I think about my agents: they learn from me, and they learn at least as much from the people I work with.

Who's teaching your agents?

If you want your team's agents to work with everyone's know-how rather than each person's alone, join at memco.ai. If you are building for your own customers, the SDKs give your product the same shared memory, and the learning-on-the-job repository has a working harness that shows agents learning from feedback end to end.