CALO

A research project, not an off-the-shelf tool

CALO stood for Cognitive Assistant that Learns and Organizes. SRI dates the project to 2003 ^[1]. It was SRI's component of DARPA's Personalized Assistant that Learns program, a five-year project with 150 million dollars behind it ^[1].

The goal SRI states is a compact description of the ambition: a cognitive assistant that could reason, learn from experience, execute commands, explain its actions, reflect on its experience and respond robustly to surprise ^[1]. Each item on that list is something ordinary software does not do. Running a command is easy; saying why a command was run, and continuing sensibly when the situation turns out differently from what was expected, is not.

Reached through IRIS

People reached the assistant through a semantic desktop called IRIS, which stood for Integrate. Relate. Infer. Share ^[2]. Treating a desktop as semantic means handling documents, messages and tasks as things that are related to one another rather than as a flat pile of separate files, so that the assistant has something to reason over.

Working through that layer, CALO could gather the documents, emails and task content needed for a meeting, suggest the next slide for a presentation, record a transcription, and track task-related status and commitment updates ^[2]. It could also automate repetitive or mundane tasks that come up during the workday ^[2].

What separated it from a conventional scheduler or a macro tool was the intended source of its behaviour. CALO was meant to learn in the wild — through observation, interaction and reasoning — rather than being preprogrammed for a fixed set of actions ^[2]. A system built that way watches what a person actually does with documents and messages, draws general conclusions from those observations, and adjusts; in principle it can take on a kind of task nobody wrote rules for. Explanation and reflection matter here for a practical reason too: someone who cannot tell why an assistant did something has little reason to hand it more of their work.

Where it sits in the line of digital assistants

SRI says it leveraged the CALO work on virtual personal assistants to create Siri, which spun out as an independent company in 2007 ^[1]. That is the clearest way to place CALO in the history of the field: a research agenda built around an office assistant that watches, learns and explains became the starting point for a widely used consumer product.

CALO was not a beginning from nothing, either. SRI describes the Open Agent Architecture, developed in the 1990s as a programming methodology for software agents, as preceding its digital assistant work ^[3]. The question there was organisational: how separately built software agents should be arranged so that they can cooperate. Any assistant that has to pull information from several sources at once depends on an answer to that question, because the work is split across components that must talk to each other.

Read together, the two threads explain a good deal about CALO's shape. One is a tradition of architecture — getting many small agent programs to cooperate; the other is a research goal — a single assistant that learns a user's working life instead of being told about it in advance. The project sits where those two meet, which is also roughly where the digital assistants that followed it sit.