The Brain-Agent Loop

The read-write cycle that makes a knowledge base compound over time. Every signal that flows through an agent should touch the wiki in both directions - read before responding, write after learning. An agent without this loop answers from stale context. An agent with it gets smarter every conversation. Source: wiki brain-agent-loop pattern

The Core Loop

Two invariants:

1. Every READ improves the response. If you answered a question without checking the wiki first, you gave a worse answer than you could have.
2. Every WRITE improves future reads. If a conversation surfaced a new connection and you didn't update the relevant pages, you created a gap that will bite later.

Brain-First Lookup Protocol

Before calling any external API, searching the web, or answering from memory:

1. qmd search "term" - keyword match, fast, always works
2. qmd query "what do we know about X" - hybrid semantic search
3. qmd get <path> - direct page read when you know the slug
4. External sources as fallback only

The wiki almost always has something relevant. External APIs fill gaps - they don't start from scratch. An agent that reaches for the web before checking its own wiki is wasting tokens and giving worse answers.

Retrieval Routing

Brain-first does not mean one tool for every memory question:

The writeback rule is the unifying invariant: runtime memory may help choose an action, but durable memory is still the wiki, skills, automations, config, logs, and agent docs.

Compiled Truth + Timeline

Every substantial wiki page should have two layers:

Above the line - compiled truth. The current best understanding, rewritten when new evidence changes the picture. If you read only this section, you know the state of play.

Below the line - timeline. Append-only evidence log. Never rewritten, only added to. Each entry has a date, source, and what happened.

The compiled truth is the answer. The timeline is the proof. The synthesis is pre-computed - unlike RAG, where the LLM re-derives knowledge from scratch every query, the wiki has already done the work.

Source Attribution

Every fact written to a wiki page should be traceable. The discipline:

• Inline [Source: ...] citations with provenance
• Source hierarchy: user's direct statements > primary sources > API data > web search
• When sources conflict, note the contradiction - don't silently pick one

This matters at scale. Six months from now, someone reads a page and can trace every claim back to where it came from.

The Compounding Thesis

From gbrain: "Most tools help you find things. A brain makes you smarter over time."

The difference between a folder of notes and a brain:

Boil the Lake

"AI-assisted coding makes the marginal cost of completeness near-zero."

When ingesting a source, do the complete thing - update all affected pages, not just the obvious one. A single source typically touches 5-15 pages. The extra effort costs seconds with an LLM. "Ship the shortcut" is legacy thinking from when human time was the bottleneck.

Three Layers of Knowledge

From gstack's Search Before Building ethos:

1. Tried and true - standard patterns, battle-tested. The risk isn't that you don't know them - it's assuming the obvious answer is right without checking.
2. New and popular - current best practices, blog posts, ecosystem trends. Search for these but scrutinize - the crowd can be wrong about new things.
3. First principles - original observations from reasoning about the specific problem. The most valuable. The best work both avoids reinventing the wheel (Layer 1) while making brilliant observations that are out of distribution (Layer 3).

Deterministic Collectors

From gbrain Skillpack Section 16: when an LLM keeps failing at a mechanical task despite repeated prompt fixes, stop fighting the LLM. Move the mechanical work to code.

┌─────────────────────────┐     ┌──────────────────────────┐
│  Deterministic Script   │────>│       LLM Agent          │
│                         │     │                          │
│  • Pull data from API   │     │  • Classify and judge    │
│  • Generate URLs/links  │     │  • Add commentary        │
│  • Track state          │     │  • Run wiki enrichment   │
│  • Output markdown      │     │  • Draft responses       │
│                         │     │                          │
│  CODE - 100% reliable   │     │  AI - judgment, context  │
└─────────────────────────┘     └──────────────────────────┘

The wiki's scripts/ directory already follows this pattern. Freshness checks, index building, and linting are deterministic scripts. The LLM handles synthesis, cross-referencing, and answering questions.

The Dream Cycle

From gbrain: the most important maintenance job runs during downtime. Periodically:

1. Sweep recent conversations for entities and concepts mentioned
2. Check each against the wiki - create or enrich thin pages
3. Fix broken cross-references, stale claims, orphan pages
4. Consolidate: ephemeral context becomes durable knowledge

In this wiki, the scripts/check-freshness.ts and weekly lint automation serve this purpose. The wiki gets smarter between sessions because the maintenance infrastructure catches what slips through during fast conversations.

The Memex Vision

Vannevar Bush described the memex in 1945: a personal knowledge store with associative trails between documents. Bush's vision was closer to this wiki than to what the web became - private, actively curated, with the connections between documents as valuable as the documents themselves.

The part Bush couldn't solve was maintenance. LLM agents handle that. The wiki stays maintained because the cost of maintenance is near zero. The human's job: curate sources, direct analysis, ask good questions, think about what it all means. The LLM's job: everything else.

Concept Position

---

Timeline

• 2026-07-01 | Concepts category refresh added this page to the Brain, memory, and retrieval family, linked it to Concept System Map, and kept it standalone because it owns this reusable mental model: The read-write cycle that makes a knowledge base compound over time. Every signal that flows through an agent should touch the wiki in both... Source: User request, 2026-07-01