Partners
Zetaris checks every quote, Meterless remembers and screens, NVIDIA Nemotron plans and finds; each doing real work you can follow into the code.
| Partner | Its job in Konkyo | Without it |
|---|---|---|
| Zetaris | The evidence store: every passage, its own reading of every PDF, and the SQL that checks every quote; topic watches as tables over OpenAlex | Nothing can be marked verified |
| Meterless | H-MEM keeps your corrections; the World Model keeps who said what; Markovian screens deep searches in bounded steps | No memory of your corrections, no record of where sources stand |
| NVIDIA Nemotron | Plans answers, embeds every passage and question, screens deep-search finds, guards what's public, and stands in for every OpenAI agent role | No search by meaning (every vector is Nemotron's) |
| ElevenLabs | Transcribes recordings to the second, records audio briefs, reads answers aloud, hears the orb | No lectures, briefs or voice |
| Composio | Gmail, Google Drive and Notion imports | Upload by hand |
| OpenAlex | Open papers for deep search and watches, and citation records | No open papers in deep search or watches |
Each also has a page on konkyo.xyz that follows one action into the code, with an example (live for Zetaris and OpenAlex): made with Zetaris, Meterless, NVIDIA Nemotron, ElevenLabs, Composio and OpenAlex.
Follow one answer
Eight steps, from the moment you ask to the receipt under the answer, each tagged with the partners doing the work and the file it lives in.
- 1
Your memory, first
Before planning, the corrections you've given Konkyo are read from your H-MEM, only the ones about the same thing as this question.
- 2
The Planner
Rewrites the question so it stands alone, plans one to three searches in the words your passages would use, and flags a comparison.
NVIDIA Nemotron ↗packages/agents/src/answer.ts - 3
The Retriever
Searches your passages by meaning and by words at once and merges the two by rank. Nothing close enough means “This isn't in your library yet.”, never a guess.
NVIDIA Nemotron ↗packages/agents/src/retrieve.ts - 4
The Writer, and a pre-check
One finding per sentence, each with a citation and an exact quote, streamed as it writes (OpenAI gpt-6-luna; Nemotron 3 Ultra stands in when it's busy). Meanwhile Zetaris confirms it holds every document those passages came from.
- 5
The Analyst, for comparisons
Beside the Writer: claims against sources, each stance with a quote checked in Zetaris, kept in your World Model.
- 6
The Verifier
Each quote in code, then all of them in one Zetaris statement against its own reading of the PDF, while a model judges whether each sentence's quotes support it.
- 7
Rewrite, or drop
Rejected sentences go back to the Writer with the reasons, at most twice. Whatever still fails is dropped; nothing unverified is shown as verified.
OpenAIpackages/agents/src/answer.ts - 8
The receipt
The steps, sources, memories used and the SQL, under the answer; the full trace in logs/<runId>.jsonl. Your 👍 or 👎 goes back to your memory.
Meterless ↗apps/web/components/chat/run-receipt.tsx
