The context layer you own
Your AI's mind
Every durable fact, where it came from, and where it can go next.
What it knows
Latest entries, with receipts
Imported history
Life signals rail #1: email
Detected by pattern matching over exports you imported — no model calls, bodies never stored. Heuristics can be wrong: every record keeps the evidence it was built from. Verify before acting.
Where this mind can go
Own every layer you can — each of these surfaces reads the same memory you see here, and none of them can keep it from you.
- Portfolio — one markdown document of everything above, pasteable into any AI tool.
- MCP — Claude Desktop/Code, Codex, Cursor, Windsurf, Gemini CLI via
sovereign mcp. - Editors & browser — VS Code, JetBrains, and browser extensions in
integrations/. - Verified export — checksummed archive from Settings, or
sovereign export --encrypt.
Knowledge base: 0 documents indexed (0 with embeddings).
Your private intelligence layer
Everything here answers to you.
Model, memory, knowledge — assembled on your machine, with receipts, portable the day you want out.
Everything you need, under your control.
SovereignAI keeps your conversations, documents, and durable memory in your own local database.
Quick workflows
Put your context to work
Readiness
Workspace health
Pick up where you left off
Recent conversations
Conversation
New conversation
Private retrieval
Knowledge base
Ground answers in documents indexed on your machine.
Knowledge
Ask your documents
Local library
Documents
Durable context
Long-term memory
Review exactly what your AI carries from one conversation to the next.
Memory you can inspect and control
Personas with memory enabled receive these notes as private context. Edit stale details or remove anything you no longer want used.
0 saved
What your AI remembers
Learned behavior · local control plane
Fine-Tuning Studio
Turn conversations you explicitly choose into reviewed training examples, then run real LoRA training on your own trainer.
SovereignAI snapshots only the conversations you choose. Starting training sends that immutable snapshot to the configured self-hosted trainer—not to OpenAI. Learned adapters can memorize private content, so review every example.
Specialize behavior without surrendering your data
Fine-tuning is best for stable behavior, response formats, and domain language. Use Knowledge instead when the goal is to teach changing facts.
Fine-tuning project
Define what should change
Describe one stable behavior that examples can teach. Keep changing facts in the Knowledge base.
Choose data and record consent
Sources are opt-in. Choose at least two independent conversations so one group can remain an untouched holdout. A review copy is created; edits never change the originals.
Review and redact every example
Exclude poor pairs, remove sensitive details, and verify the ideal assistant answer. High-risk flags must be acknowledged when the dataset is locked.
Splits are grouped by conversation so near-identical turns cannot leak into evaluation. Training requires at least two independent groups; smaller snapshots remain exportable.
Run actual training on your trainer
SovereignAI uploads only the approved train and evaluation JSONL blobs. The optimizer and GPU dependencies live outside this Node process.
Training controlsGuided defaults
Approve a dataset and verify the trainer before starting.
Evaluate before deployment
Trainer metrics and the untouched holdout are evidence, not a guarantee. Record an explicit decision before assigning the model to a persona.
Deploy with lineage intact
The trainer must attest to an already-registered Ollama model and its digest. SovereignAI verifies that digest before assigning the model to a persona; it never sends the training dataset to Ollama.
Workspace controls
Settings
Configure where intelligence runs and what context it can use.
General
Name your workspace and choose its default intelligence.
Providers
Choose a local runtime or connect an API you control.
Personas
Shape how your AI behaves and which context each role can access.
Model Studio
Design, save, export, and build model recipes you control on your configured Ollama endpoint.
Recipes stay in this workspace. Build artifacts are stored by the Ollama endpoint in your saved provider settings.
Building does not train or fine-tune model weights. If selected, quantization creates a derived lower-precision artifact. The base model's license still applies.Memory behavior
Control whether conversations can automatically create durable notes.
When set, memory-writing calls (auto-learning, distillation) always run this model on your default provider — a small local model can own what gets learned about you while chat uses anything. See the starter shelf in Model Studio for picks.
Knowledge embeddings
Semantic retrieval improves meaning-based search. Keyword search remains available without it.
Pull locally with ollama pull nomic-embed-text.
Data & privacy
Your SQLite database stays in the SovereignAI state directory. Exports copy workspace data; chats may send prompts and selected memory or knowledge to the configured model provider.
The customs declaration: the system prompt, memories, knowledge excerpts, and prior messages that would leave, with sizes, before anything does. Local endpoints never ask — nothing leaves. Every answer from a remote provider carries a receipt of how much left and to which host.
Export personas, conversations, memories, and document chunks to one readable JSON file. API keys are never included.
Bring conversations from ChatGPT's or Claude's official data export (or a generic JSON shape for anything else — Grok, Kimi, GLM, DeepSeek, Qwen, etc.) into this workspace. Parsing happens locally; nothing is uploaded anywhere else. Re-importing the same file is safe — already-imported conversations are skipped, not duplicated.