Canonical Distiller Engine
Transforms messy multi-turn transcripts, voice notes, and chat archives into typed nodes (Decision, Idea, Knowledge, Artifact, Process) while automatically linking contradictions and superseding outdated decisions.
Universal memory for every AI agent. Lojin SuperBrain V3 captures scattered insights from AI coding sessions, voice notes, Obsidian, Slack, and web clips — distills them via a 6-agent council into a cited knowledge graph, and compiles them into live MCP skills your agents execute on demand.
Compiling recurring engineering fixes into MCP skills cuts onboarding time by 70% across all coding agents.
Unverified web clippings must never auto-compile into skills. Enforce minimum provenance trust = AI_SESSION.
Proceed with mandatory human-in-the-loop approval and provenance trust gating before publishing to MCP.
"Why did we switch our static file serving and how is GPU reranking configured?"
Canonical containment verification (Path.resolve().is_relative_to()) was mandated to eliminate CWE-22 path traversal [DEC-2026-014]. Stage-2 GPU reranking uses nvidia/llama-3.2-nv-rerankqa-1b-v2 over RRF candidates with automatic local fallback [KNW-2026-089].
Most tools dump raw text into a vector database and hope for the best. SuperBrain runs a coordinated council of 6 specialized AI agents that structure, debate, verify, and compile your institutional memory.
Transforms messy multi-turn transcripts, voice notes, and chat archives into typed nodes (Decision, Idea, Knowledge, Artifact, Process) while automatically linking contradictions and superseding outdated decisions.
Answers natural-language queries across Personal, Department, and Company vaults. Every claim carries mandatory inline citations to source nodes — and explicitly abstains when the vault lacks verified evidence.
Scores every captured idea across Feasibility, Market Impact, Novelty, Viability, and Alignment, then convenes a 3-persona adversarial debate (Optimist, Skeptic, Pragmatist) with an Arbiter verdict.
Compiles validated engineering and operational know-how into structured SKILL.md artifacts. Tracks real-world execution telemetry from coding agents and auto-revises skills when edge cases fail.
Sweeps the graph every night to merge semantic duplicates, link orphan nodes, flag stale decisions, and generate an Executive Morning Briefing summarizing 24-hour organizational velocity.
Logs every capture, distillation, BYOK verification, and MCP tool invocation in an append-only ledger with token counts, latency telemetry, and per-tenant USD cost attribution.
Run 100% locally on SQLite + multi-provider BYOK (Gemini, OpenAI, Anthropic), or toggle optional NVIDIA NIM & NeMo Retriever microservices for TensorRT-LLM inference and Stage-2 Cross-Encoder reranking.
Combines dense semantic vectors, exact lexical keyword matching, and 2-hop graph topology traversal into a unified Reciprocal Rank Fusion score.
Plug in an NVIDIA API key or self-hosted NIM container to unlock TensorRT-LLM completion and Stage-2 Cross-Encoder neural reranking with zero lock-in.
Every node is scoped across 3 organizational boundaries and 4 cryptographic provenance trust tiers so unverified web clips never pollute core truth.
Static documentation rots the moment it is written, and generic RAG chatbots invent answers without accountability. SuperBrain closes the loop from capture to agent execution.
| Capability | Notion / Passive Wikis | Standard Vector RAG | Lojin SuperBrain V3 |
|---|---|---|---|
| Knowledge Capture | ✕ Manual copy-paste & formatting | ✕ Static PDF / doc uploads only | ✓ Live MCP, Voice, Slack, Obsidian & Chat Archives |
| Contradiction & Drift Handling | ✕ Outdated pages silently coexist | ✕ Retrieves conflicting chunks blindly | ✓ Typed supersedes & contradicts graph edges |
| Search & Reranking Precision | ✕ Basic keyword search | ✕ Single-signal cosine similarity | ✓ 3-Signal RRF + NVIDIA NeMo Cross-Encoder Rerank |
| Strategic Idea Evaluation | ✕ None | ✕ Generic chat opinion | ✓ CEO 5D Radar + 3-Persona Adversarial Debate |
| Executable Agent Skills | ✕ Humans must read & re-prompt | ✕ Passive Q&A only | ✓ Auto-compiled SKILL.md served live via 15 MCP Tools |
| Provenance & Cost Governance | ✕ Page edit history only | ✕ Uncited black-box generations | ✓ 4 Trust Tiers + Append-Only Secretary Cost Ledger |
Three movements. Capture becomes typed structure. Structure becomes strategic judgment. Judgment becomes executable skills your AI agents run automatically.
One-click register the MCP server into Claude Code, Antigravity, or Cursor. Link your Obsidian vault, Telegram voice bot, Slack workspace, or drag-and-drop full ChatGPT / Claude / Gemini export archives.
The Distiller extracts decisions and SOPs; the CEO Agent scores ideas in 5D and runs multi-persona debates; the Nightly Curator deduplicates, auto-links graph edges, and prepares your Morning Briefing.
Query cited answers mid-session from any AI IDE, promote validated workflows into the Skill Factory, and let your agents load approved SKILL.md playbooks on demand. Every run feeds telemetry back into the brain.
Everything you need to know about deployment, local-first privacy, MCP compatibility, and optional NVIDIA GPU acceleration.
Leave your details and our engineering team will set up a tailored walkthrough with your own workflow, vaults, and AI agents.
Or email us directly at superbrain@lojin.id