The continuing loop

HALO uses the terms of the survey Always-On Personal Intelligence. This page gives the definitions the plugin relies on. Read mode uses the same definitions, written out in halo/read/protocol/PROTOCOL.md.

Three modules

An always-on personal assistant runs a loop of three functional modules.

• Sensing (S) acquires and interprets current observations: user input, sensor streams, screen state, external events, and the observed state of the assistant’s own execution.
• Memory (M) retains information across interactions.
• Action (A) selects the assistant’s behavior: silence, communication, preparation, or execution.

Operations are assigned by function, not by component name. One model, tool, or shared context can play several roles, so an agent that reads a calendar, keeps a preference, and sends a reminder plays all three.

Six channels

A coupling channel is a directed information flow in which one module supplies information that another module uses for its decisions or state updates.

Channel	Information that flows
S→A	current input reaches Action. Every agent system has this basic path.
S→M	sensed or user-supplied evidence is admitted to Memory.
M→A	retained information guides an Action decision: timing, content, effector, commitment, or silence.
A→S	Action guides Sensing: what to observe next, or a check of whether an action had its expected effect.
M→S	retained priors guide Sensing: what to attend to, or how to interpret new input.
A→M	action context, records, or outcomes are kept in Memory.

In the plugin, a channel X→Y appears when an event of module Y used an event of module X. Event model shows how agents state that.

Five evidence steps

An assistant that improves itself changes its own mechanism: a prompt, rule, skill, policy, or model. The survey asks for a chain of evidence that runs from the assistant’s own records to a revision it later uses. HALO judges each step on its own.

1. Record. The assistant keeps records of its own actions for later use.
2. Outcome. Later evidence about those actions, such as a reply, a rating, a correction, or a task result, is obtained and kept.
3. Attribution. An outcome is linked to the decision it concerns and used to check or correct assistance. An immediate check counts even if nothing is kept.
4. Revision. The assistant itself makes a durable change to its mechanism. Facts and preferences used by a fixed mechanism are content, not a revision, and changes written by users or developers are not revisions by the assistant.
5. Later use. Later steps run under the revised mechanism.

Channel coverage and the evidence chain are separate dimensions. The survey found that no reviewed work reporting every channel also reports a revision, and that works with a complete chain leave some channels unspecified. HALO therefore reports them in separate tables and does not combine them into a score.