Overview

Living documentation for the Knowledge Layer — how the principal's knowledge is captured, labeled, deduplicated, superseded, and reasoned from. Every number and rule here is pulled from the live code, not from memory.

The whole loop in one picture

There are two completely separate flows, and they never write to the same place at the same time:

• The WRITE side (ingestion). A single background worker reads the team's conversations once on a schedule, turns conversation into a small, deduplicated model of how the principal thinks, and is the only process allowed to write the corpus.
• The READ side (answering). A read-only API. When the agent layer gets a question, it calls one endpoint — /retrieve — and gets back matching knowledge plus the principal's reasoning patterns. It never writes knowledge.

Why split this way: the two jobs have opposite needs. Writing wants care, dedup, and LLM judgment (slow, once a day). Reading wants to be fast and never blocked. Keeping one writer also means two processes never fight over the same file or row — no write races. This is the single-writer rule.

The five ideas to lead with

If you read nothing else, read these. They are the design decisions everything else follows from.

1. It models reasoning, not quotes. Every entry must pass one test — "would this help answer a question the principal was never directly asked?" That test is what lets the system speak on new topics instead of just replaying old ones.

2. Fewer, stronger entries. A meeting yields a handful of entries, deduplicated — not a transcript. Repetition strengthens one entry instead of creating ten copies. (An earlier design produced 270 entries from a single meeting; this one is built to do the opposite.)

3. Two honesty dials. Confidence (how sure) and stability (how durable) are judged separately. So core principles never decay in ranking, while current takes are allowed to.

4. Context is first-class. When the principal runs the same reasoning move the opposite way in a different setting, that is stored as a variant with its context — not flagged as a contradiction.

5. It changes its mind when the principal does. A new opposing position supersedes the old one (which is kept in history), so the model tracks how thinking evolves over time.

How to read the rest of this site

The pages are ordered from concept to reference:

• Capturing knowledge — how raw conversation becomes a few durable entries.
• Anatomy of an entry — every field on a knowledge entry and how its value is computed.
• The living model — dedup, corroboration, and supersession: why the corpus is a model, not a log.
• Reasoning patterns — the "snowball" that powers extrapolation.
• The answer path — how a question becomes a ranked response.
• API reference — the read-only contract used by the agent layer.
• Configuration — every tuning knob, its value, and why.
• Operations & reliability — how the worker stays correct and self-heals.