Three beliefs about a personal brain (operator-held)

Operator-held opinions designed for the takes bootstrapper. Concrete, falsifiable, high-stakes. Joint-relevant because both operators run the same substrate and face the same design decisions.

The cost matrix presented at install is a menu but the operator's actual decision is a binary — does the agent treat its budget as a measurable constraint, or as plenty? Cost × scale curves matter more per-query than they matter by absolute spend. For a personal brain under 10K pages, the per-query token cost is in the noise; the marginal recall is worth it.

Without it, dream / agent / autopilot (the LLM-bearing subagent features) all refuse to submit. The take-bootstrap, fact extraction, and pattern synthesis phases stall. gbrain calls this "the cross-section where compounding happens" — and getting it wrong leaves you with a search box dressed as a brain.

Tested with the gbrain corpus — quantum-entanglement vs. unrelated sentences scored 0.84 vs. 0.38 from a Q8_0 quant of Qwen3-Embedding-0.6B on a 1 vCPU VPS. Sufficient discrimination for a small brain (≤ 10K pages). At 100K+ pages the recall ceiling matters more.

How it's structured

Belief 1: A local embedder on a CPU VPS gives 95% the retrieval quality of a hosted embedder Tested with the gbrain corpus — quantum-entanglement vs. unrelated sentences scored 0.84 vs. 0.38 from a Q8_0 quant of Qwen3-Embedding-0.6B…
Belief 1: A local embedder on a CPU VPS gives 95% the retrieval quality of a hosted embedder Risks: low-traffic domains, multilingual corpora (training was English-heavy), very long chunks.
Belief 2: tokenmax is the right default search mode for a single-user brain, regardless of model The cost matrix presented at install is a menu but the operator's actual decision is a binary — does the agent treat its budget as a measura…
Belief 2: tokenmax is the right default search mode for a single-user brain, regardless of model Risks: 100K+ pages where tokenmax crosses real spend limits.
Belief 3: ANTHROPIC_API_KEY (or equivalent subagent route) is required, not optional, for a "compounding" brain Without it, dream / agent / autopilot (the LLM-bearing subagent features) all refuse to submit. The take-bootstrap, fact extraction, and pat…
Belief 3: ANTHROPIC_API_KEY (or equivalent subagent route) is required, not optional, for a "compounding" brain Risks: vendor lock-in, key rotation discipline.
… 2 more sections in the full essay