Chat locally. Keep your context. Use the model you choose.
Explore features · Quick start · Architecture · Build from source · Docs
Made by Cyanex · Local inference by default · Network features opt-in
ADAM is a local-first personal AI application, not just a chat window around a model. It can stream local responses, save and resume conversations, remember and correct personal context, answer questions about supported local documents with source references, and work with approved models on your device.
With additional setup, ADAM also supports voice, OCR, Android browser access, opted-in web research and bounded computer actions with explicit review. The key design decision is that memory, knowledge, identity and permissions belong to ADAM, not to the model weights. Switching models does not mean starting over.
| 🧠 Memory with history Keep, correct, trace or forget facts |
📚 Knowledge with sources Ask your own local documents |
🔁 Replaceable models Choose from approved GGUF models |
🔐 Explicit control Opt in to network and review actions |
| Capability | What ADAM can do | |
|---|---|---|
| 💬 | Local AI chat | Run a local language model, stream replies and resume saved chats. Optional Careful mode uses two passes to review the question and answer. |
| 🧠 | Persistent memory | Save preferences and facts, correct stale information, inspect provenance and forget information on request. Memory persists across model switches. |
| 📚 | Document Q&A | Import supported local files and answer with source references. Keyword search works without a separate embedding model. |
| 🔗 | Personal context | Explore evidence-linked relationships, timelines, goals and on-demand briefings. Inferred routines are labeled suggestions. |
| 📦 | Local model library | Scan supported folders for GGUF models, approve and switch models, or use selection based on task complexity and available memory. |
| ⚙️ | Performance controls | Choose CPU or Vulkan GPU inference, toggle Eco / Priority and inspect model-loading state, queues and resources. |
| 🎙️ | Voice | Transcribe locally, review speech before sending and read responses aloud. Conversation mode and speaker recognition need extra setup. |
| 👁️ | Image / window text | OCR images or a specifically selected Windows application window. General image understanding needs a configured vision model. |
| ⏰ | Reminders | Schedule reminders, receive them in the inbox and use a visible background tray indicator. |
| 🌐 | Internet research | Opt in to supported search providers, fetch public HTTPS pages and use a text browser. Regular chat stays local by default. |
| 📱 | Android access | Pair an Android browser over the local network. The phone client has fewer permissions than the desktop owner. |
| 🔐 | Profiles & privacy | Separate owner, guest and fictional demo profiles, with owner authentication and credential vault support. |
| 🩺 | Diagnostics & recovery | Inspect service, model, resource and storage health; export privacy-scoped diagnostics; create encrypted backups, restore in stages and use supported update rollback. |
| 🧪 | Behavior experiments | Review synthetic examples, export a training dataset and recipe, and compare evaluation reports. Training itself is a separate process. |
| 🛠️ | Reviewed computer actions | Approve supported tool operations and bounded See → Plan → Act → Verify workflows on supported Win32 controls. |
Note
This table describes the application's capabilities, not a promise that every one is active in the offline starter bundle. Some features require separately installed models, supported hardware, the native desktop app, provider configuration or explicit approval. See interfaces and extra setup.
|
Keep conversations, memories and evidence in ADAM's application layer. Correct a preference, trace where it came from, or forget it. Then switch between approved models without moving your personal state into the model itself. |
Search supported local documents and follow source references. Keyword retrieval works even without an optional semantic embedding model. Linked personal context is grounded in supporting memories rather than treated as unquestionable fact. |
|
Optional speech, OCR, reminders, model controls, research and paired Android access make ADAM useful in different contexts. Capabilities are gated by their interface, permissions and installed components. |
Ordinary inference stays on-device by default. Internet access is opt-in. Mobile is permission-scoped. Supported computer actions require review; model output alone does not get to operate your computer. |
Use ADAM-0.20.0-Source-With-Model.zip for the preconfigured source terminal experience on Windows x64.
- Extract the complete ZIP to a writable folder.
- Double-click
RUN_ADAM.bat. - Chat locally. Enter
/quitto exit and shut down the model server started by the launcher.
For a separate profile with fictional project data, launch RUN_DEMO.bat instead.
Important
The full offline bundle includes Python, llama.cpp and a verified Qwen3-4B Q4_K_M chat model. It needs no global Python install or pip download. The first launch checks the model checksum. The React/Tauri desktop application is a separate build; the ZIP starts the source terminal.
What's inside the bundle?
ADAM/
├── RUN_ADAM.bat One-click editable source terminal
├── RUN_DEMO.bat Fictional demo profile
├── backend/adam/ Python application source
├── apps/desktop/ React + Tauri desktop source
├── models/ Chat model, checksum and upstream notices
├── runtime/
│ ├── python/ Bundled interpreter and dependencies
│ └── llama/ Local inference runtime
└── data/ Private state, created on first use
The bundled CPU profile uses 3 threads, a 2,048-token context and up to 384 output tokens. Performance depends on your machine and other running applications. Speech, speaker, semantic-search and vision models are not part of the default chat-model bundle.
/remember preference editor = Example Studio
/correct editor = Another Editor
/show memory --history editor
/why know editor
/forget editor
The fictional demo also includes example project context:
/graph Cedar
/timeline Cedar
/goals
/briefing
The GitHub checkout/source ZIP does not include runtime binaries, model weights or personal data. See build from source to install dependencies and the model.
flowchart TB
subgraph Clients["INTERFACES"]
direction LR
D["Desktop / React + Tauri"]
T["Source terminal"]
M["Paired Android browser"]
end
M --> G["Pairing + scoped LAN gateway"]
D --> A["Authenticated local API"]
G --> A
A --> B["Profile + permission boundaries"]
B --> Core["ADAM runtime"]
T --> Core
Core --> Context["Identity, memory, local knowledge"]
Core --> Models["Model provider + resource checks"]
Models --> Engine["llama.cpp / configured adapters"]
B --> Personal["Relationships, timelines, goals"]
Personal --> Context
B --> Actions["Reviewed actions + reminders"]
B --> Recovery["Encrypted backups, restore, rollback"]
classDef entry fill:#142b43,stroke:#8396ff,color:#f0f5ff
classDef core fill:#08333d,stroke:#2dd4bf,color:#ebfffd
classDef state fill:#143327,stroke:#6ee7b7,color:#ebfff2
classDef protected fill:#302442,stroke:#c4b5fd,color:#f5f3ff
class D,T,M entry
class A,G,B,Core,Models,Engine core
class Context,Personal state
class Actions,Recovery protected
The terminal is a trusted local entry point. Desktop and Android requests pass through authentication and profile boundaries. Model-generated text does not itself authorize computer operations; separate action providers check permissions and state. See the architecture specification and threat model.
flowchart LR
P["Approved local model A"] --> R["Selected inference provider"]
Q["Approved local model B"] -.->|Switch models| R
R <--> A["ADAM application layer"]
A <--> S[("Memory · documents · identity · permissions")]
classDef model fill:#1d2545,stroke:#8b84ff,color:#ffffff
classDef core fill:#0e3240,stroke:#20ceca,color:#ffffff
classDef data fill:#153a2e,stroke:#6ee7b7,color:#ffffff
class P,Q model
class A,R core
class S data
The model processes requests. ADAM keeps the user's state. Model selection can change without migrating stored personal context into a new set of weights.
flowchart LR
S["Memory + source evidence"] --> P["Profile + memory policy"]
P --> K["Retrieve with provenance"]
K --> A["Answer / personal view"]
S --> C["Correct or retract"]
C --> H["Keep history; invalidate stale links"]
S --> F["Forget"]
F --> R["Remove dependent links and goals"]
classDef stored fill:#0e3240,stroke:#20ceca,color:#ffffff
classDef output fill:#1d2545,stroke:#8b84ff,color:#ffffff
classDef change fill:#302442,stroke:#c4b5fd,color:#ffffff
class S,P stored
class K,A output
class C,H,F,R change
Relationships point back to supporting memories rather than overriding evidence. Historical facts remain distinguishable from current facts; routines are suggestions, not established facts. Briefings are requested manually, and derived personal context follows the same correction and forgetting rules.
| Component | What to know |
|---|---|
| Offline starter bundle | Provides the source terminal, Python, llama.cpp and default verified chat model. |
| Desktop UI | React + Tauri requires its native setup/build. A frontend dev preview alone is not the native desktop application. |
| Voice / speaker features | Optional speech and speaker assets must be configured separately; conversation mode and speaker recognition require additional setup. |
| OCR / vision | OCR extracts text from a chosen image or window. Full image interpretation requires a separate vision model. |
| Semantic search | Optional; local document keyword retrieval works without embeddings. |
| Vulkan GPU | Depends on supported hardware and Vulkan-capable runtime configuration; CPU is supported. |
| Internet research | Requires opted-in provider configuration; ordinary chat remains local by default. |
| Android browser | Requires pairing and a scoped local-network gateway; its permissions differ from the desktop owner's. |
| Reviewed Win32 actions | Only supported controls/workflows; operations are bounded and require explicit review. |
| Behavior training | Synthetic dataset generation, recipe export and evaluation comparisons are available; actual LoRA/QLoRA training is separate and not validated as an included fine-tuned adapter. |
The demo profile contains fictional projects and a scoped note sandbox for practicing write, verify and undo without owner data.
Requires Windows x64 and Python 3.13.
# From the repository root: install the locked dev environment.
.\scripts\dev.ps1 -Install -Command init
# Download and verify the pinned model and inference runtime.
.\scripts\setup-model.ps1
# Launch the source terminal with CPU inference.
.\scripts\start.ps1 -CpuThe helper creates its environment under %LOCALAPPDATA%\ADAM\dev-env. Alternatively, use requirements.lock.txt and python -m pip install -e . in a manually managed environment.
Frontend development
cd apps/desktop
npm ci
npm run devThe dev server is only a frontend preview. Follow the desktop setup guide for the native bridge and builds.
The v0.20.0 release records the following checks:
| Check | Recorded result |
|---|---|
| Python regression suite | 463 passed |
| Frontend / native Rust tests | 12 / 3 passed |
| Live source / frozen service checks | 48 / 48 passed |
| Scripted framework + security evaluation | 22 cases, 0 critical failures |
| Offline source bundle | Real Qwen CPU test; fresh ZIP extraction and BAT startup passed |
Caution
These are recorded release results, not live CI status. The 22-case benchmark tests scripted framework and security probes, not general LLM quality. VRAM and voice measurements were unavailable; human acceptance, clean-VM and alternate-hardware validation are tracked separately.
Run the checks yourself
Use the project's locked development environment:
python -m pytest -q
python -m ruff check backend tests
python -m mypy backend/adam
python -m adam.evals.runner --suite full --summaryThe evaluation runner uses scripted transport unless real-model mode is selected. More detail: evaluation methodology.
Copyright © 2026 Cyanex. ADAM is source-available, licensed under the Community Non-Commercial License v1.0. Commercial use requires written permission or a separate commercial license. Review contribution terms before contributing; third-party models and runtimes have their own terms in THIRD_PARTY.md.
Commercial licensing: [email protected] · Cyanex on GitHub
ADAM · Local intelligence. Persistent context. Explicit control.
Built to keep your context with you, even when the model changes.