Nate Jones Second Brain

When intelligence is abundant, context becomes the scarce resource. This skill is context architecture — a persistent, searchable knowledge layer that turns your agent into a personal knowledge manager.

Two opinionated primitives:

• - Supabase — your database, and so much more. PostgreSQL + pgvector. Stores thoughts, people, projects, ideas, and tasks as structured data with vector embeddings. REST API built in. Your data, your infrastructure. Models come and go; your context persists. And once you have a Supabase project, you've unlocked the foundation for everything else you'll want to build — the Second Brain is just the beginning.
• OpenRouter — your AI gateway. One API key, every model. Embeddings and LLM calls for classification and routing. Swap models by changing a string. Future-proof by design.

Everything else — how you capture thoughts, how you retrieve them, what you build on top — is application layer. The skill covers the foundation.

If the tables don't exist yet, see INLINECODE0

Building Blocks

These are the operational concepts behind the system. Understanding them helps you operate correctly.

Full conceptual framework: INLINECODE7

Five Tables

Every table has semantic search via its own match_* function. Cross-table search via search_all.

Routing Rules

When a thought is classified:

If confidence < 0.6, don't route. Leave in thoughts, tell user.

Quick Start

Capture a thought (full pipeline)

CODEBLOCK0

Full pipeline with routing logic: INLINECODE23

Semantic search (single table)

CODEBLOCK1

Cross-table search

CODEBLOCK2

Returns table_name, record_id, label, detail, similarity, created_at from all tables.

List active projects

CODEBLOCK3

List pending tasks

CODEBLOCK4

Ingest Pipeline

When content arrives from any source:

1. 1. Embed the text via OpenRouter (1536-dim vector)
2. Classify via OpenRouter LLM (type, topics, people, confidence, suggested route)
3. Log in thoughts (the Receipt — always, regardless of routing)
4. Bounce check — if confidence < 0.6, stop here
5. Route to structured table based on type (the Sorter)
6. Confirm to the user what was captured and where it was filed

Full pipeline details: INLINECODE31

Metadata Schema

Every thought gets classified with:

References

• - Conceptual framework: INLINECODE50
• First-time setup: INLINECODE51
• Database schema (SQL): INLINECODE52
• Ingest pipeline details: INLINECODE53
• Retrieval operations: INLINECODE54
• OpenRouter API patterns: INLINECODE55

Env Vars

Security Notes

Why service_role key? Supabase provides two keys: anon (public, respects RLS) and service_role (full access, bypasses RLS). This skill uses service_role because:

• - This is a single-user personal knowledge base, not a multi-tenant app
• Your agent IS the trusted server-side component
• The RLS policy restricts access to service_role only — the most restrictive option
• Using the anon key would require loosening RLS to allow anonymous access to your thoughts, which is worse

Data sent to OpenRouter: All captured text (thoughts, names, action items) is sent to OpenRouter for embedding and classification. This is inherent to the design — you need AI to understand meaning. Don't capture highly sensitive information unless you accept OpenRouter's data handling policies.

Key handling: Store SUPABASE_SERVICE_ROLE_KEY and OPENROUTER_API_KEY securely. Never commit them to public repos. Rotate periodically. In OpenClaw, store them in openclaw.json under skills.entries or as environment variables.