Lobu

Agents that scrape a snapshot and forget it are useless for account health, contract review, or variance analysis — the signal lives in the delta between yesterday and today. Lobu addresses this by maintaining a persistent, structured memory layer tied to live connectors, so agents scan for change rather than reprocess everything from scratch. The core workflow is four steps: connect data sources, define a watcher goal with an approval condition, let the agent scan on schedule and attach evidence to memory entities, then review and act on what surfaces in your team’s chat or via the API.

The differentiating design choice is the watcher-plus-approval model. Instead of agents that fire actions automatically, Lobu lets you specify exactly when the agent must pause and surface a draft for a human to edit, send, or discard. The vendor describes this as the agent ‘keeping evidence attached’ — meaning the reasoning behind a churn flag or contract risk isn’t buried in a log, it’s stored alongside the memory entity so the person reviewing it can audit the signal before acting.

Lobu fits teams building multi-user agents where shared organizational memory matters: a revenue agent that multiple CSMs query, a legal agent that routes contract risk to the right reviewer, a finance agent that prepares variance reports on a schedule. It runs locally via a single CLI command, self-hosted via Docker or Helm for teams where data cannot leave their cloud, or as a managed cloud offering. The open-source codebase and 13 public example workflows on GitHub give teams a concrete starting point rather than a blank canvas.

Where it breaks: watcher definitions are goal-oriented and work well for monitoring and drafting workflows, but the docs describe no built-in visual branching or conditional routing between agents. Teams that need one agent’s output to trigger a different agent down a decision tree will find themselves writing that logic in the connector SDK — a separate code layer that needs its own testing and maintenance. At that point the simplicity of the watcher model is mostly gone, and teams running complex multi-step orchestration often evaluate purpose-built agent frameworks instead.