Lojin SuperBrain
The company brain for your AI agents.
SuperBrain captures decisions from AI chats, Slack, voice notes and docs, distills them into a typed knowledge graph, and serves that context to Claude Code, Cursor, Gemini CLI and any MCP agent, with citations.
Working product, used daily to run MIDAZ. Self-hosted today; cloud workspace in progress.
SuperBrain, running on our own vault.
The full 3-minute walkthrough is shared privately with design partners and investors.
- distilled knowledge nodes
- 477
- distilled knowledge nodes
- compiled agent skills
- 90
- compiled agent skills
- MCP tools
- 15
- MCP tools
AI has amnesia. Teams pay for it every day.
Knowledge evaporates
Decisions made in an AI chat, a Slack thread or a voice note disappear the moment the session closes.
Transcripts are mostly noise
Exploration and abandoned options dominate raw transcripts. Feeding them to agents burns tokens and degrades answers.
Search is not execution
Document search tells an agent about your company. It does not tell the agent how your company works.
One closed loop from raw conversation to executable skill.
- 1
Capture
MCP, Slack, Telegram voice, a Chrome extension, Obsidian and ChatGPT / Claude / Gemini exports feed one ingestion queue. SHA-256 hashing rejects duplicates.
- 2
Distill
A canonical distiller keeps only durable primitives: decisions, ideas, knowledge and artifacts. A typical session yields 1 to 3 nodes, not 10.
- 3
Graph
Nodes link through a closed set of typed edges such as supersedes, depends_on, contradicts and duplicate_of. Conflicts go to human review.
- 4
Reason
The Librarian answers with mandatory citations and declines when evidence is weak. The CEO agent pressure-tests plans against past decisions.
- 5
Execute
The Skill Factory compiles approved decisions into SKILL.md files that agents discover and run over MCP.
Built for agents. Governed for teams.
Native MCP server
Fifteen tools for capture, recall, insight versioning, graph relations and skills, callable mid-session from any MCP-compatible agent.
Hybrid retrieval
Vector similarity and full-text search fused by reciprocal rank fusion, over distilled nodes and the raw transcripts behind them.
Zero-regression vault
Raw sources are kept permanently, so the whole archive can be re-distilled as models improve. Supersede edges and soft deletes only.
Secretary audit ledger
An append-only record of every capture, query and review, with per-call model, token and cost accounting, enforced by database triggers.
Nightly curator
Re-embeds, de-duplicates, auto-links and clusters themes overnight, then delivers a morning briefing of what changed.
Bring your own keys
Per-tenant encrypted API keys. Gemini by default, with optional NVIDIA NIM inference and NeMo Retriever embeddings and reranking.
# tools/list
◆ superbrain_capture◆ superbrain_recall◆ superbrain_save_insight◆ superbrain_update_insight◆ superbrain_supersede◆ superbrain_get_node◆ superbrain_list_nodes◆ superbrain_verify_node◆ superbrain_explore_relations◆ superbrain_confirm_relation◆ superbrain_reject_relation◆ superbrain_list_skills◆ superbrain_get_skill◆ superbrain_report_skill_run◆ superbrain_log_actionAvailable inside Claude Code, Cursor, Gemini CLI and any MCP host.
Numbers from our own benchmarks.
- local vector search, down from 15.0 ms after matrix optimisation (19× faster)
- 0.8 ms
- local vector search, down from 15.0 ms after matrix optimisation (19× faster)
- recall@10 retained after pruning embeddings from 3,072 to 1,536 dimensions
- 96.2%
- recall@10 retained after pruning embeddings from 3,072 to 1,536 dimensions
- vector storage after dimensional pruning
- −36%
- vector storage after dimensional pruning
- automated tests across API routes, concurrency, tenant isolation and crypto
- 350+
- automated tests across API routes, concurrency, tenant isolation and crypto
Internal benchmarks on a single-node deployment of our working vault. Methodology available on request.
Works where your team already works.
Agents
- Claude Code
- Cursor
- Gemini CLI
- Any MCP client
Capture
- Slack
- Telegram voice
- Chrome extension
- Obsidian
- ChatGPT export
- Claude export
- Gemini / Takeout
Models
- Gemini 3.5 Flash-Lite
- Gemini Embedding 2
- NVIDIA NIM (optional)
- NeMo Retriever (optional)
Stack
- Python · FastAPI
- FastMCP
- SQLite + vector index
- Next.js
- PostgreSQL + pgvector (cloud, in progress)
Also on the SuperBrain core
Lojin Ops V2
Operational governance for hybrid human and AI teams: proof-of-work checks, heartbeats and cost telemetry.
Lojin Connect
An AI relationship manager for WhatsApp that answers, qualifies and books with company context from SuperBrain.
Programs & certifications
Member of NVIDIA Inception Program. Google Cloud & AI certifications held by our founder.
Where we are.
- Phase 1 Shipped
Capture & memory core
Omnichannel capture, MCP server, distiller, vault.
- Phase 2 Shipped
Graph & intelligence
Temporal graph explorer, Librarian and CEO agents, nightly curator, Skill Factory.
- Phase 3 In progress
Cloud workspace
Multi-tenant team workspaces, PostgreSQL + pgvector, SSO.
- Phase 4 Next
Autonomous action loop
Outbound execution agents with explicit permission classes.
Built by an operator who needed it first.
Samuel Frans
Founder & AI Systems Architect
Samuel has built production systems across fintech (payment gateways and POS), media automation and AI orchestration. He designed SuperBrain’s zero-regression, multi-agent governance architecture and runs every MIDAZ venture on it daily.
LinkedIn- Company
- MIDAZ Dev
- HQ
- Jakarta, Indonesia
- Program
- Member of NVIDIA Inception
- Contact
- [email protected]
We are onboarding a small group of design partners.
For AI-native teams of 10 to 250 people who use Claude Code, Cursor or other agents every day. You get hands-on onboarding and a direct line into the roadmap. We get honest feedback.
Investor or ecosystem partner? Write to [email protected] for the deck and a live walkthrough.