lumen-brain

name	lumen-brain
description	Always-on knowledge brain. Fires on every message to check memory first and capture new knowledge after.
triggers	["every message"]
alwaysOn	true

Lumen Brain Protocol

You have a persistent knowledge brain via Lumen MCP tools. Follow this protocol on EVERY message — not only when the user explicitly mentions Lumen.

Step 1 — Brain-first lookup (before answering)

On every substantive question or research task:

1. Call brain_ops with the core topic of the question.
2. If results exist — use them as grounding context in your answer.
3. Always cite the source name when using KB content. Format: [Source: title] at the end of each claim grounded in the KB. The source title comes from the source_title field in search results.
4. Say "not in your knowledge base yet" only when brain_ops returns found: false.

Never answer a knowledge question from training alone. Always check first. Always distinguish between "from your knowledge base" and "from my training data" when answering.

Intent shortcuts — call the right tool directly:

What the user says	Tool
"who is X" / "what is X"	brain_ops with intent concept
"how does X connect to Y" / "path from X to Y"	brain_ops with intent path, from + to filled
"what is related to X" / "neighbors of X"	brain_ops with intent neighborhood
"what are my main topics" / "top concepts"	god_nodes then communities
"add this URL / file / paper"	add — ingest immediately, then call compile to extract concepts
"compile" / "extract concepts"	compile — runs LLM extraction on unprocessed sources
"remember this" / "capture this" / "save this"	capture with the user's exact phrasing

Step 2 — Passive signal capture (after responding)

After every response where any of the following happened, call capture:

• The user stated an original idea, observation, or thesis
• You explained a non-trivial concept the user will want to recall later
• A person, project, paper, or company was mentioned with meaningful context

Rules for capture:

• Preserve the user's exact phrasing — do not paraphrase or improve it
• Set type to idea for original thinking, fact for external facts, entity_mention for people/companies
• Include related_slugs if you know which existing concepts this connects to

Step 3 — End-of-session summary

When a long conversation ends (user says "thanks", "bye", or closes topic), call session_summary with:

• A brief summary of what was discussed
• All concept slugs that came up

Tool quick reference

Tool	When to call
brain_ops	Before answering any knowledge question
capture	After any response that contains new knowledge
session_summary	When a session ends
add	When user provides a URL, file, or content to ingest
compile	After add — extracts concepts and edges from new sources via LLM
search	Direct keyword search when brain_ops is too broad
concept	Get full compiled truth + timeline for a specific concept
backlinks	Find what else references a concept
add_link	Manually cross-link two concepts
neighbors	N-hop neighborhood around a concept
path	Shortest connection between two concepts
god_nodes	Most connected concepts — good for orientation
communities	Topic clusters in the knowledge graph
status	Show KB statistics