Skip to content

Memory

A mind that remembers.

The library is what the office has. Memory is how it finds things and remembers what matters.

A wide museum-exhibit visualization on a clean white background showing how an agent's memory is structured — pinned facts, entities, dated incidents, stories that connect them, and one final story loaded at every cold-start.
Memory of an agent — the final story rebuilt daily, loaded at every cold-start.

At the top level, memory knows where everything is. Below that, memory knows what’s inside each folder. Below that, memory knows the stories — the threads of meaning that connect a hundred files. A story might say “this customer has been unhappy for three months across these eight emails.” That story has to be findable. It also has to be matchable against new data as it arrives. When a new email lands, memory asks: “does this connect to a story I already know?” Today memory is a combination of vector search (for fuzzy matching) and filesystem index (for exact retrieval). The big unsolved problem is how memory should be structured when an office has been running for years.

How memory is built

Memory grows from notes taken during work. Notes cluster into stories. Stories cluster into one final story — the cold-start brief the agent reads on its way in every morning.

A wide museum-exhibit visualization on a clean white background showing how memory is built: notes → stories → one final story loaded at every cold-start.
From notes to stories to one final memory — built every day from what the office has actually done.

Vendors we track

top-level indexfolder searchstory matchingvector + filesystem hybridcold-start inject

Going deeper

🚧 Sub-pages are coming online. For the complete six-layer picture, see the six-layers overview.