Your confidential AI. Already on it.
Local-first encrypted memory — one chat room, Digital Hats for specialists, and an MCP synapse that plugs into the tools you already use.
Download · Docs · Ghampus · Changelog · Un-Brain paper · Borrowable Skills
Cloud chat memory is a sticky note on someone else’s server. Graphnosis is a dual-graph cortex on a folder you control: typed edges, temporal confidence, skills-as-SOPs, and deterministic recall you can audit. Ghampus orchestrates one room; Agempi wear Digital Hats and join when you @ them — work lands in living Topics, not a pile of dead chat threads.
| Typical AI “memory” | Graphnosis | |
|---|---|---|
| Where it lives | Vendor cloud | Your disk (or your server) |
| Structure | Flat notes / chat crumbs | Dual-graph engrams + edges |
| Procedures | Prompt again | Walkable Skills (SOPs) |
| Who answers | One chat persona | Ghampus + hatted Agempi |
| Clients | One product | Claude, Cursor, Zed, Cline, phone, HTTP MCP… |
| Forgetting | Opaque | Soft-delete, tiers, consent gates |
Un-Brain navigation — Your AI / Ghampus, Cortex, Synapses, Engrams, Agempi, and the rest of the product surface.
An Agempus is a specialist persona (avatar, role, skills). Several together are Agempi. Each wears a Digital Hat — visual shorthand for who is speaking.
Real early Digital Hats from dogfooding Graphnosis — the repo author’s first Agempi, trained as living specialists on the same cortex the product builds.
Empty Ghampus room — the orchestrator explains how to start and when to bring specialists in.
| Idea | What it means |
|---|---|
| Ghampus orchestrates | Untagged messages stay with Ghampus. He routes, watches memory, and can offer Save to your cortex? |
@ agents / @everyone |
Invite by name, crew, or a ranked set — not a free-for-all flood |
| 1:1 Focus | DM-like specialist lane in the same room; mentions pause; Ghampus still watches for memory saves |
| Remember / Reply / Quote / Comment | In-window actions — approve memory writes before they land |
| Topics vs Chat Sessions | Topics = living work lanes. Chat Sessions = soft-archived transcripts after you clear a topic. Cortex memory is untouched by clear. |
Ask Ghampus how agents collaborate — skills and subgraphs stay in the room.
@mention a Digital Hat when you want a specialist lane — Ghampus still orchestrates the room.
Guide: Ghampus chat — topics, hats & one room. Story: Ghampus.
Graphnosis is built to meet you where you already work — not to trap you in one chat box.
| Surface | What you get |
|---|---|
| AI clients (MCP) | Claude Desktop, Claude Code, Cursor, Zed, Cline, Kimi, Copilot bridge, and any MCP-aware client — same synapse, sensitivity tiers, consent gate |
| In-app Ghampus | One room + Digital Hats on desktop and phone UI |
| Connectors | RSS, GitHub, Slack, Linear, Obsidian, webhooks, and more — credentials stay on-device |
| Mobile & remote | Browser UI + HTTP/SSE MCP bridge; LAN or Tailscale; QR unlock |
| Self-host | Headless Linux/Docker personal server — cortex encrypted on your disk |
| Skills / packs | Train SOPs, export signed .gsk, borrow a skill another Agempus already has instead of redoing the work |
Connect AI → getting-started/connect-ai · Phone → getting-started/mobile · MCP tools → reference/mcp-tools
- Personal — researchers, writers, builders who want AI that recalls their notes, decisions, and files across sessions.
- Developers — Cursor / Claude Code / VS Code users who want repo context, runbooks, and walkable SOPs on the MCP wire.
- Teams — shared runbooks and onboarding skills without re-pasting every week (Job Memory Kits).
- Air-gapped / regulated — memory stays on-device; optional local LLM (Ollama); no cloud back-end for the cortex.
- Download Graphnosis for macOS / Windows / Linux.
- Unlock a cortex folder (local disk or synced op-log via iCloud/Drive).
- Open Ghampus, create or enable Agempi, connect an AI client via Copy MCP Config.
Full guide: deploy/linux/README.md. Shape of the path:
# On the Linux host (build on Linux — native addons are platform-specific)
git clone https://github.com/nehloo-interactive/GraphnosisApp.git
cd GraphnosisApp
pnpm install
pnpm -r build
GRAPHNOSIS_CORTEX=~/graphnosis-cortex \
GRAPHNOSIS_PASSPHRASE='your-passphrase' \
GRAPHNOSIS_HTTP_UI=1 \
GRAPHNOSIS_BIND=0.0.0.0 \
GRAPHNOSIS_HTTP_UI_PORT=3456 \
GRAPHNOSIS_HTTP_UI_TOKEN="$(openssl rand -hex 32)" \
GRAPHNOSIS_HTTP_UI_STATIC="$PWD/apps/desktop/dist" \
node apps/desktop-sidecar/dist/index.jsThen open http://<host>:3456, unlock with the token, and (recommended) put it behind Tailscale Serve. Systemd unit + Docker notes live in the same deploy guide. Mobile walkthrough: Connect from your phone.
Prerequisites: Node 20+, pnpm 10+, Rust (desktop shell), Bun (Tauri build compiles sidecar/MCP binaries via bun build --compile).
git clone https://github.com/nehloo-interactive/GraphnosisApp.git
cd GraphnosisApp
pnpm install
pnpm -r build
pnpm --filter @graphnosis-app/desktop-sidecar smoke # encryption → ingest → recall → forget
pnpm dev:desktop # menu-bar app (first compile is slow)Standalone sidecar:
GRAPHNOSIS_CORTEX="$HOME/Graphnosis" \
GRAPHNOSIS_PASSPHRASE="dev-passphrase-change-me" \
pnpm dev:sidecarRemote-cortex dev tooling (laptop UI → host sidecar): see deploy/launcher/README.md.
Tauri shell (Rust) ── Unix socket JSON-RPC ── Node sidecar (TypeScript)
│ │
│ menu bar, keychain, Share Sheet │ cortex, ingest, MCP, Skills
│ │ connectors, mobile / HTTP UI
└──────────────────────────────────────────────┴── @nehloo/graphnosis SDK
Engine: @nehloo/graphnosis (Apache 2.0) · Sync: graphnosis-secure-sync (FSL-1.1-Apache-2.0).
Even if you’re “just” auditing or filing issues, this is the interesting surface:
- Dual-graph cortex — directed + undirected edges over the same nodes; temporal scoring; deterministic recall.
- Skills as SOPs — compile, walk, export signed
.gskpacks; Agempi borrow a skill another hat already owns instead of reinventing it. - Local-first everything — encryption, op-log, consent gate, sensitivity tiers (
public→deidentified→sensitive). - One-room UX — Topics, Focus,
@routing, Remember cards — product model inapps/docs/.../ghampus-chat.md. - Connectivity — MCP tool surface, connectors, mobile bridge, headless deploy.
Good first issue themes (file against Issues — no fake ticket numbers):
- Docs clarity (self-host, MCP client setup, Topics vs Sessions)
- Connector edge cases / new connector sketches
- Accessibility & mobile layout polish
- Eval fixtures for recall honesty / contradiction triage
- Packaging / deploy scripts for air-gapped installs
See CONTRIBUTING.md for the current engagement model: docs PRs and issues are welcome; security reports go to security@graphnosis.com. Code PRs are not open yet while the encryption / consent / op-log paths are still under adversarial review — the source is published so you can read, audit, and self-host.
Two preprints so far — declarative memory first, then procedures on the same graph. Short versions below; Zenodo for the citable PDFs, nehloo/graphnosis-whitepaper for sources and figures.
The Un-Brain: A local-first, indelible knowledge multi-graph for private AI memory with deterministic, owner-adjudicated contradiction handling · DOI 10.5281/zenodo.20843387
What it is: the memory substrate behind Graphnosis — a co-equal dual graph (directed + undirected over the same nodes), soft-delete / indelibility, and owner-adjudicated contradictions (no silent LLM “fixes”). Evaluated on LongMemEval.
Why care: if you’re auditing recall honesty, temporal confidence, or “why not flat RAG,” this is the formal argument and the numbers.
Engram connectivity in 3D — the dual-graph cortex you can actually see.
Borrowable Skills as Lean Un-Ganglia Subgraphs · DOI 10.5281/zenodo.21205599
What it is: Skills as SOPs compiled into typed-edge subgraphs in the same owner-held cortex as declarative memory — not a separate skill store. Sibling Agempi borrow a skill by name instead of spawning a whole agent; packs travel as signed .gsk. Continues The Un-Brain.
Why care: if you care about walkable SOPs, Digital Hats borrowing each other’s procedures, or routing / token cost of skill packs, this is the paper.
Code Skills / Ship Workflow as a walkable subgraph in the same Engram cortex as declarative memory — borrowable SOPs, not a separate skill store.
Functional Source License 1.1, Apache 2.0 Future License (FSL-1.1-Apache-2.0).
You may read, audit, fork, modify, and self-host — including inside your organization for any internal purpose (commercial internal use included). Consulting, training, and custom integrations for organizations using it internally are expressly permitted. No seat count or revenue threshold applies to running it yourself.
You may not (during the two-year window) offer Graphnosis — or a modified version — to third parties as a hosted or embedded product/service that substitutes for it (no Graphnosis-as-a-Service).
Each published version becomes available under Apache License 2.0 on the second anniversary of the date that version was made available. See LICENSE → Grant of Future License.
Self-host stays free and full for privacy. Paid tiers add product capability for teams that prefer not to self-host.
| Product | graphnosis.com |
| Download | graphnosis.com/download |
| Docs | Getting started |
| Ghampus | graphnosis.com/ghampus |
| Chat model | Guides → Ghampus chat |
| Skills | Reference → Skills |
| Un-Brain | DOI 10.5281/zenodo.20843387 · repo |
| Borrowable Skills | DOI 10.5281/zenodo.21205599 |
| Changelog | v1.37.0+ |
Made by Nehloo Interactive LLC.