Compiled truth + timeline

GigaBrain pages have a peculiar shape: a top half that is always current and a bottom half that is append-only. This isn’t just a markdown convention — it’s the central design choice that makes a personal brain coherent over years.

The split

Every page renders, in order:

1. Frontmatter — typed metadata (title, type, slug, tags, links).
2. Compiled truth — the body above the line. Edited in place. Treated as the canonical, always-current statement.
3. A horizontal rule — the literal --- separator.
4. Timeline — append-only entries below the line. Treated as evidence.

---
title: Alice Example
type: person
---

# Alice Example

Founder at RiverAI. Works on offline-first knowledge systems.
Met through a shared project in Q1 2026; reliably thoughtful in
asynchronous threads.

---

## Timeline

- 2026-04-14 — Met at a demo day. Interested in offline-first knowledge.
- 2026-04-22 — Replied to outreach; warm intro to Bob.
- 2026-05-03 — Co-authored a draft on local embeddings; sharp edits.

The compiled-truth section answers “what is true about Alice today?” The timeline answers “how do we know?”

Why two halves

A single notebook entry can either describe the world now or record what happened. Most note systems force you to pick:

• A wiki page is always-current — but you lose the audit trail.
• A daily journal is append-only — but you lose the synthesized view.
• A messy mix of both, and you lose both.

The compiled-truth + timeline split is borrowed from Andrej Karpathy’s compiled-knowledge model: above the line is the model’s parameters, below the line is the training data. When new evidence arrives, you append to the timeline and re-compile the top.

The benefits compound:

• Always-current summary at the top. Search results and snippet displays use this; you don’t get stale fragments.
• Audit trail at the bottom. Provenance, dates, sources — all preserved.
• Cheap synthesis. Re-compiling truth is editing one paragraph, not rewriting the whole page.
• Honest uncertainty. When the timeline disagrees, contradiction detection catches it; you don’t have to.

What this means for retrieval

Hybrid search treats the two halves differently:

• The compiled-truth section is chunked by heading and embedded as a coherent unit. Vector search returns these.
• Each timeline entry is embedded individually. Date-anchored questions (“what happened in April?”) return the relevant entries directly.

The result: a query about a person’s current title returns the compiled section; a query about when something happened returns timeline rows. You don’t have to design two indexes — the page model gives you both for free.

What this means for writing

The first time you write a page, you write the timeline. The compiled truth comes later, when you have something to compile.

A typical lifecycle:

1. Capture. Drop a timeline-style line: - 2026-04-14 — Met Alice at a demo day. Offline-first knowledge.
2. Accumulate. Two or three more entries arrive over weeks.
3. Compile. Edit the top of the page to summarize: “Alice founded RiverAI; works on offline-first knowledge.”
4. Repeat. Each new piece of evidence appends to the timeline; the compiled truth gets edited in place when needed.

gbrain timeline-add and the brain_put MCP tool both make this easy. You never write to the timeline by hand if you don’t want to.

What this means for assertions

Contradiction detection (brain_check) reads triples out of the body and the frontmatter. When the compiled truth says “Alice is at RiverAI” and a recent timeline entry says “Alice left RiverAI,” the assertion machinery flags the mismatch.

This is why we keep both halves: the timeline is the corpus the assertion engine compares against, and the compiled truth is what it audits.

What this isn’t

• Not version control. Git is fine for that. The timeline is editorial; rolling back the compiled truth is just an edit.
• Not a journal. Journal pages exist (type: journal) but they’re a special case — the entire body is a timeline entry.
• Not a graph. The graph is in the links table. The page model is about the content shape, not the structure between pages.

Summary

Two halves, one page:

• Above the line: what is true now. Edited in place.
• Below the line: how we know. Append-only.

Everything else — search, contradiction detection, ingestion, ergonomics — falls out of that split.