konkyomade with back to konkyo

memory, the world model and long tasks

Tell it once.
It remembers.

Mark an answer wrong and say why: Meterless H-MEM keeps that correction, with where it came from, and the next answer reads it first. Its World Model keeps where each of your sources stands on each claim.

step into the library No account needed: walk in as a guest. Your memory is on the knows page.
interactive examplemeterless × konkyo
Step
you asked

Does a daytime nap help motor learning?

what was wrong?
Not in the sourceWrong sourceMissed somethingNot what I asked

“say which group each number is from”

The path one correction takes, in the app's own words: the memory H-MEM stores and the note the next answer's Planner and Writer read.

Follow one correction.

From one small action to what you see,
each step opened in the code.

  1. 1

    You mark an answer wrong.

    A 👎 with a reason, a cited passage that didn't fit, or a line you tell it on the knows page.

  2. 2

    H-MEM keeps it, with its source.

    The correction goes into your own H-MEM with its provenance: which answer, which chat, and how much to trust it (0.8 for a 👎, 0.9 for a line you wrote).

  3. 3

    The next answer reads it first.

    Before planning, H-MEM ranks your memories for the new question, and only those about the same thing are kept.

  4. 4

    The Planner and Writer follow it.

    The relevant corrections go into their instructions as guidance, never as evidence: a memory is never cited.

Meterless in Konkyo

Where it
shows up.

Specific jobs in the app.
Each one has a place you can point to.

What it knows, in one place.

The knows page lists how you like answers, from H-MEM. Cross one out and it's forgotten; H-MEM keeps the deletion in its trust ledger.

Where your sources stand.

When an answer compares sources, the Analyst writes each source's stance on each claim into a World Model: agrees, disagrees or partly, with the quote it was read from.

Spaces learn together.

Each person has a World Model, and so does each shared space: it grows from everyone's asks in it, and its members see it on the space's knows page.

Deep search in bounded steps.

Markovian screens what deep search found eight abstracts at a time, carrying only the shortlist, so the context stays the same size however many papers come back.

A closer look.

The choices behind the behaviour.
Open whichever part you're curious about.

Why recall is re-ranked

H-MEM's built-in embedder hashes words, so “memory” in two unrelated questions looks alike to it. Konkyo keeps H-MEM's ranking, then drops anything under 0.35 cosine on Nemotron embeddings: related questions measured 0.54–0.68, unrelated ones 0.00–0.20.

Feedback adjusts trust

A 👍 or 👎 on an answer reaches the memories that shaped it, through H-MEM's own feedback, so a correction that keeps helping gains weight.

Read back, never made up

The knows page reads claims from the rows the Analyst wrote, never from a model. The World Model refuses a write without provenance.

Vendored as published

H-MEM, the World Model and Markovian live in packages/meterless, unchanged from upstream commit 0aa7417 (Apache-2.0), each with its own file storage.