Brains

Origin

Every conversation with a language-model assistant starts at zero. Voice preferences drift, project state is re-explained, hard-won corrections evaporate. After enough sessions you realize you are not actually working with an assistant — you are conducting a series of first dates that happen to be productive.

Problem

What's needed is not a longer context window. What's needed is structure: a place where facts, directives, corrections, decisions, and predictions live as first-class rows that any assistant in the rotation can read and write. The cognitive infrastructure has to outlive the conversation.

Approach

A Postgres schema for first-class persistent cognition. Tables: facts, directives (rules to follow), corrections (mistakes not to repeat), decisions (with alternatives_considered and beliefs_relied_on), predictions (with verification windows), voice_laws (how to write), tool_profiles, mirrors. Operates as an MCP server; every assistant in the rotation reads from and writes to the same brain. Predictions are graded against reality on a twelve-hour loop; calibration drift triggers wisdom-note updates across the directive set.

Methodology

The schema is opinionated by design. There is no 'notes' bag. A correction has a wrong_value, a correct_value, a danger_level, and a context — because the kind of correction that matters has structure, and the kind that doesn't, doesn't belong. A decision has alternatives_considered because the alternatives are usually what makes the decision interesting. Voice laws are versioned because what one assistant calls 'tone' another calls 'formatting' and the only way to keep them in sync is to write the rules down where both can read.

Selected milestones

• Operational across multiple assistants
• Twelve-hour prediction-verification loop in production
• Voice laws + corrections enforced on every session start

Open questions

• How calibration drift propagates across distinct domains (markets vs. conservation vs. infrastructure)
• Where the line is between a correction (rule-shaped) and a fact (claim-shaped)
• What happens when two assistants disagree about a directive's scope

Ask me about

• How predictions are graded against reality on the twelve-hour loop
• Why the schema is so opinionated — and what gets thrown out
• What 'voice laws' are, structurally, and how they survive across model upgrades
• How a correction differs from a fact in the schema

Stack