Announcement · May 16, 2026 · 4 min read
Introducing Memory in Company Agents.
Agents that forget every job cannot run a business. Memory is the layer that lets them remember what matters, and forget what does not.
Memory is live in Company Agents. It is the layer that holds what Agents have learned, what humans have decided, and what each company’s particular preferences are. Three layers, each tuned for a different lifetime.
Three layers
Short-term scratch lives only for one Issue. It holds the working state of one agent doing one job. When the Issue closes, the scratch is gone. Project memory persists for the lifetime of a project. It holds the decisions that were made, the brand voice that was chosen, the gotchas that were discovered. Shared memory sits at the company level. Every Agent in the company can reach it. It holds the things that are true about how the company works, not about how any one project ran.
If the next Agent needs it, write it to Memory. If only this Agent needs it, keep it in scratch.
The rough rule
How writes work
Memory is event-sourced. Every write is an append, with a source (which Agent, which Issue, when), a topic, and a body. Nothing is overwritten. You can read the history of any topic to see how a decision was arrived at, or how the brand voice has drifted over time.
How reads work
Reads happen through MCP Tools that Memory exposes. An Agent does not pull all of Memory into context; it queries Memory for the topic it needs. The read is shallow by default and can be deepened with follow-up queries. This keeps context windows tight.
What ships today
All three layers are live. The Memory Tools are available to every Agent that loads the matching Skill. The Memory inspector in the UI lets humans browse what is in any layer for any project or company.