Meterless
Three Meterless engines in working features, your corrections, who said what, and deep searches in bounded steps.
The engines are vendored under Apache-2.0 in packages/meterless and driven from
packages/agents.
H-MEM: your corrections
A not right with a reason, a ๐ on Right passage?, a practice question you mark This is
wrong, or anything you tell it becomes a memory, with where it came from. The Planner reads
the ones that matter before every answer, and the Tutor before it writes practice. ๐ and ๐ on an
answer adjust the confidence of the memories it used. You see them all on
What Konkyo knows; cross one out and it's forgotten. (packages/agents/src/memory.ts)
World Model: who said what
Each comparison is recorded as sources, claims and supports / contradicts / qualifies
relations, each with the passage it was read from, in your World Model and, when you asked in one
space, the space's. That's where your sources stand on the knows page. (packages/agents/src/analyst.ts)
Markovian: long searches in bounded steps
Deep search screens what it found in steps of eight, carrying only the shortlist forward, instead
of one long prompt. Its own report of tokens saved is shown and traced (markovian_run): on testing it screened 38 sources in 5 steps, 6,569 tokens used, about 12,472 saved, labelled as an
estimate. (packages/agents/src/scout.ts) How deep search works