How is this different from an experiment tracker?

Experiment trackers log metrics and parameters for humans to compare. Glen is memory for agents: they retrieve the relevant prior findings and write new ones automatically in one MCP call, so context surfaces while the agent is actually writing or analyzing code.

Does the whole ML team share one memory?

Yes. Glen is org-shared, so every ML engineer's agents read and write the same store. A negative result or data quirk one person discovers is immediately available to everyone else's agents.

Shared memory for ML engineers

Machine learning engineers accumulate a quiet mountain of context: which feature transformation broke the pipeline, why a model was rolled back, the hyperparameter range that never helps, the data quirk that explains a weird eval result. Most of it lives in experiment trackers no one revisits, in notebooks that rot, and in the memory of whoever ran the experiment, so the team relearns dead ends and re-debugs the same data issues. When ML engineers bring in AI agents, to write training code, analyze runs, or debug pipelines, those agents have none of that history and happily suggest the approach that already failed. Glen, shared memory for AI agents, gives those agents a durable, team-shared memory as a single MCP tool that retrieves relevant prior context and records new findings in one round trip.

Connect Glen to the agents your ML team uses, a coding agent writing training and eval code, a notebook assistant, an experiment-analysis helper, and each gets one tool that reads the relevant history and writes back what it discovered. Before an agent proposes an approach, it pulls what the team already knows: the architecture that underperformed, the data-leakage bug that inflated a past metric, the preprocessing step that matters for this dataset. After a run, it records what worked, what failed, and why, so the next experiment builds on it instead of repeating it. You stop relitigating settled questions and stop losing hard-won negative results.

The shift is that experimental knowledge becomes a shared, durable asset rather than scattered across individual runs and people. Glen is org-scoped, so the memory spans every ML engineer and every agent on the team, and even downstream agents that consume the models. One engineer's agent records that a certain feature leaks the label; every other agent avoids it next time. Because Glen is a standard MCP server, the same memory is readable from any MCP client, so a training-code agent, a data-debugging agent, and an eval analyst all draw on one consistent record of what the team has learned. Wire it in once over OAuth or an API key and let your experimental knowledge, especially the failures worth remembering, compound across projects.

FAQ

How is this different from an experiment tracker?: Experiment trackers log metrics and parameters for humans to compare. Glen is memory for agents: they retrieve the relevant prior findings and write new ones automatically in one MCP call, so context surfaces while the agent is actually writing or analyzing code.
Does the whole ML team share one memory?: Yes. Glen is org-shared, so every ML engineer's agents read and write the same store. A negative result or data quirk one person discovers is immediately available to everyone else's agents.