What's New in Cornelius v10.26

Released 2026-10-08 · Cornelius versions are named for their month (vMM.YY)

Corneliusis the open-source second-brain agent built for Trinity — a company or personal brain that compounds, with receipts. This is a Cornelius release, not a Trinity one. New here? Start at Cornelius.

v10.26 is about reasoning you can check. Cornelius now decides, not just advises; every conclusion runs the same battery of pre-conclusion checks; every read declares why it reads; and similarity scores are stated against the population they run in. Plus a graph fix that needs one rebuild after upgrading, and a template for what to trust at ingestion.

What changed

•/decide — decisions, not just advice — when the question is “X or Y?”, “go/no-go”, or “is it worth it”, /decide expands the real option set (status quo, defer, pilot, hybrids), classifies the decision (reversible or one-way door, one-shot or repeated, ruin exposure, risk vs radical uncertainty), applies the matching rule (ergodic filter first, then expected value, robustness / minimax regret, or value of information), and closes with tripwires. A recommendation with no named rule is advice wearing a costume.

•Reasoning checks — a shared discipline contract (reasoning-checks) that /advise, /decide, and crystallization run before concluding: an epistemic inversion (pre-mortem with a concrete falsifier), an attractor check (is this conclusion just the gravity of your most-linked notes?), provenance weighting, and a reference-class anchor for any estimate.

•Read roles (scope-mount) — every knowledge-base read now declares why it reads: voice reads (your perspective, your articles) see only your own thinking; reasoning and lookup reads also see what you have read, grouped as your thinking · what you have read · records. Reference scopes (records, per-book shelves) mount only when the question names them. No more silent reads of the wrong slice of the vault.

•Similarity scores you can trust — SIMILARITY-CALIBRATION.md: a score means nothing without its population, so every threshold in the system is now stated against the space it actually runs in. Search results carry raw_similarity next to the learning-adjusted ranking score, and calibrate_similarity.py re-measures the bands on your vault.

•Graph fix — rebuild your index — identical YAML frontmatter blocks made unrelated notes look like perfect twins (cosine 1.000), crowding real neighbours out of the semantic graph. index_brain.pynow skips frontmatter-only and empty chunks when choosing a note's representative; tension.py uses the same rule.

•Source authority — SOURCE-AUTHORITY.md is one template for what to trust at ingestion: a tier model (primary · credible-interpreter · discovery-tier · rejected), a web allowlist, and reject patterns. discovery-tier lets an off-list source raise a question but never settle one.

•Skills ship with their scripts — the book extractors (multi-format-book-extractor, epub-chapter-extractor) and benchmark-memory now include the Python they call.

•60 skills — plus refreshed advise, recall, find-connections, dialectic, ingest-source, create-article, refresh-index, and the extraction agents. See the skills catalog.

After upgrading

The graph fix changes how notes are represented in the index. Rebuild it and re-run tension detection:

cd resources/local-brain-search && ./run_index.sh
# then, in Claude Code
/detect-tensions

Learn more

How Cornelius works · Getting started · All skills

Source and full history: github.com/Abilityai/cornelius

All releases