Metsu — The Living Cognitive Distillation Engine

Everyone is racing to give AI memory of your facts and preferences. I’m building the layer underneath. Metsu is a model of how you actually think: your beliefs, your mental models, your reasoning style, how you decide, and how you make sense of your own life. Grounded in validated psychology, cited to your own words. It ports into any AI to make it yours, and you can read it to understand yourself.

Metsu (滅): a Buddhist word for cessation, the end of suffering. Also a backronym, Meaning Extraction Thought Synthesis Unit.

What it is, plainly

A pipeline that reads your own raw output (notes, voice memos, chat logs) and distills a structured model of how you think. Not your facts. Not your preferences. Your cognition. The model is portable, so you drop it into any LLM, and readable, so you can review it yourself.

Two faces, one engine

A, the promise: optimize any AI for you. A machine-readable cognitive profile that personalizes any model far past “remembers your tone.” This is the headline, and the larger need in an AI-saturated world.
B, the depth: a mirror you can read. The same model rendered for a human. It surfaces patterns in how you think and narrate your own life that you might not see yourself. This is the uncontested, scientifically richest layer, and it is what makes A credible.

Why it’s different

The field treats personalization as a storage problem: remember more facts. It is actually a cognitive-modeling problem. Metsu grounds personalization in validated cognitive science and models the layer nobody else does, the life-story layer, with every claim cited to your own words. Depth is the moat, not memory size.

The rules it won’t break

This is the trust story, and it’s non-negotiable.

Grounded. Every extracted claim is quoted back to its source. No uncited inference.
Two layers, kept separate. What you expressed (cited) versus what is cautiously inferred (hedged, dimensional).
No pseudoscience. No personality types, no diagnoses, no population comparisons. Metsu openly shows what it refuses to infer, and why.
Built on peer-reviewed psycholinguistics, not vibes.
Local-first. Your corpus and your model are yours. Nothing personal is committed or phoned home. This is the leg the incumbents structurally can’t copy, because their business is centralized data.

Where it is right now

In active development. The repo is private during the build; the code opens at launch, alongside the synthesis layer and the proof harness. I won’t overclaim, so here is the honest state.

Built and tested: the cognitive-science grounding research, a live competitive-landscape scan, an evidence-graded catalog of which cognitive facets are validly extractable, the core data schema with the two-layer epistemics enforced in the type system, a deterministic psycholinguistic feature layer (pure code, unit-tested), a pluggable LLM transport, the grounded extraction prompt, ingestion adapters, and the parallel extraction engine that turns documents into grounded cognitive atoms. Full test suite green.

Next: the synthesis layer that assembles the model over time, the portable output (context.md), the narrative-identity depth, and a faithfulness harness that scores groundedness. The open research question I’m working now: how to rigorously prove a personalized model produces better outputs.

Follow it

The research behind Metsu (the cognitive-science grounding, the facet catalog, the personalization-proof protocol) stays private while the engine is being built. It’s the moat. metsu.ai → is where the project itself lives and where it’ll surface when there’s something to show.

The builds it grew from

Voice Distillation →. Turns voice-memo transcripts into a structured second brain. MAP then REDUCE on Claude, routed through the local claude CLI, no API key. Built to run on my own recordings.
Personal Knowledge Distillation →. The earlier notes version, built on my own 700-plus notes.