OpenBrain
A personal knowledge database architecture by Nate B Jones: Postgres + PGVector + Supabase + MCP. Gives every AI tool you use persistent, searchable memory across sessions. Thousands of community members have built their own versions. Total cost: $0.10–0.30/month.
The Problem It Solves
Claude’s memory doesn’t know what you told ChatGPT. ChatGPT’s memory doesn’t follow you into Cursor. Every platform has built a walled garden of memory — five separate piles of sticky notes on five separate desks. Memory is designed as lock-in.
“Your knowledge should not be a hostage to any single platform.”
Architecture
Capture
Type a thought (Slack, any messaging app) → Supabase edge function runs in parallel:
- Generates a vector embedding (mathematical representation of meaning)
- Extracts metadata (people, topics, type, action items)
- Stores both in Postgres with PGVector
- Replies in thread with confirmation
Round trip: under 10 seconds.
Retrieval
MCP server connects to any compatible AI client. Three tools:
- Semantic search — find thoughts by meaning, not keywords. “What was I thinking about career changes?” finds notes about consulting even if “career” was never used.
- List recent — browse what you captured this week
- Stats — see your thinking patterns
Works from Claude, Claude Code, ChatGPT, Cursor, VS Code — anywhere MCP is supported. One brain, every AI.
Cost
~$0.10–0.30/month on Supabase free tier for ~20 thoughts/day in API calls. “You’ll spend more on coffee this morning.”
Setup
~45 minutes, no coding. Tested with a non-technical user who completed the full setup.
Four Lifecycle Prompts
- Memory migration — extract everything your existing AI tools know about you (Claude memory, ChatGPT memory) into OpenBrain. Run once.
- OpenBrain Spark — interview prompt discovering how the system fits your workflow; generates personalized capture suggestions by category
- Quick capture templates — 5-sentence starters optimized for clean metadata extraction: decision capture, person note, insight, meeting debrief
- Weekly review — end-of-week synthesis: clusters by topic, scans for unresolved actions, detects cross-day patterns, finds connections you missed. ~5 min on Friday afternoon.
The Compounding Advantage
Person A opens Claude, spends 4 min explaining context. Gets a good answer.
Person B opens Claude — it already knows her role, projects, constraints, team, last week’s decisions via MCP. She asks one question and gets an answer informed by 6 months of accumulated context. If she switches to ChatGPT, same brain, same quality.
Every thought captured makes the next search smarter. The gap compounds weekly.
“Memory architecture determines agent capabilities much more than model selection does.”
The Three-Primitive Stack
OpenBrain is the memory layer in Nate’s agent stack:
| Primitive | Implementation |
|---|---|
| Memory | OpenBrain (Postgres + PGVector + MCP) |
| Proactivity | loop (Anthropic-native) |
| Tools | Any MCP-connected service |
Together, these three replicate most OpenClaw capabilities without the security concerns.
Human Web vs Agent Web
The internet is forking: human web (fonts, layouts, reading) vs agent web (APIs, structured data, machine-to-machine). Note-taking apps (Notion, Apple Notes, Evernote) were built for the human web. OpenBrain is infrastructure for the agent web — your thoughts stored in a format anything can query.
Design Principles
- You own it: No platform lock-in. Postgres, not proprietary format.
- Boring stack: Postgres isn’t VC-backed, isn’t deprecating, isn’t chasing a growth metric.
- Standards-based: MCP for connectivity. No proprietary orchestration layer.
- Both capture and retrieval: Any MCP client is both a write point and a search tool.
- Community-driven: Open recipes for energy tracking, sales, networking, content, job search.
Compared to Other Memory Approaches
| Approach | Storage | Search | Auto-improvement |
|---|---|---|---|
| OpenBrain | Postgres + PGVector | Semantic (vector) + recency | No (manual) |
| Open Viking | File system | Tiered loading | Yes (auto-compression) |
| Hermes Agent | Built-in | Internal | Yes (closed learning loop) |
| Platform memory (Claude, ChatGPT) | Proprietary | Per-platform only | No (siloed) |
See Also
- loop — the proactivity primitive that gives OpenBrain a heartbeat
- OpenClaw — the framework OpenBrain + /loop replaces
- Four Prompting Disciplines — the framework OpenBrain serves (context engineering layer)
- RAG vs Wiki — OpenBrain uses semantic search; wiki uses structural search
- Open Viking — alternative file-system-based agent memory
- Nate B Jones — creator
- Source: The $0.10 System
- loop + OpenBrain