Arcanum

One engine for your AI work

Move from a question to work getting done.

Arcanum runs as a local service that connects models, tools, conversations, and memory. Use it from the terminal, connect an API client, or work through The Forge as it becomes available. Each connects to the same engine and the same durable records.

Choose the models that fit

Use a local model, a hosted provider, or both.

Connect OpenAI-compatible providers, including Ollama through its /v1 endpoint. Arcanum can also use Claude Code and Codex CLIs you have already installed and signed in to for text completions. Its own agent tool loop runs through providers that support tool calls.

Arcanum exposes an OpenAI Chat Completions compatibility subset so compatible clients can connect to it. The models, endpoints, and capabilities you configure determine what is available.

Work through the steps

Give the assistant more than a reply box.

Keep useful work moving: The agent loop can call models and tools repeatedly, use the results, and correct course. It stops on a final response, cancellation, repeated lack of progress, or a budget, context, or safety boundary.
Understand the workspace · Eye of the World: A concise view of directory structure, filenames, and recent changes helps the assistant recognize the project it is working in. Content search and retrieval provide the deeper detail.
Use tools and delegate work: Built-in tools can read, search, and edit workspace files; MCP connects additional tools. Delegate bounded tasks, or use Apprentices for longer work with plans, checkpoints, and pause and resume controls.
Reuse and evaluate your approach: Organize work into Campaigns and conversations into Sessions. Save repeatable workflows as Spells, then use Trials and their checks, called Inquisitors, to compare outputs and test behavior.

A memory system with distinct jobs

Remember useful things without losing where they came from.

Project facts, standing preferences, and lessons from a conversation need different treatment. Arcanum gives each a place, keeps its source attached, and lets you inspect and curate what is remembered.

Keep the record · Grimoire and Session history: The Grimoire is Arcanum's encrypted local store. It holds conversations and durable records, while Session history preserves the order of exchanges and the files shared with them.
Keep confirmed context distinct · Covenant: Record standing preferences and claims for you or a Campaign. Context you confirm stays separate from material the assistant proposes; a suggestion cannot confirm itself or grant tool access.
Recall specific facts · Lexicon: Keep structured information about people, projects, services, and other named subjects. Arcanum can find and update those facts with their source and scope attached.
Learn from past conversations · Saga: When enabled, Saga extracts associations and conclusions from completed conversations and recalls relevant observations later. These remain derived memories with traceable sources.
Trace a memory's history · Annals: Track the origin, version, scope, and timing of remembered claims. The Annals records evidence about memory history so a change can be understood in context.
Review and correct what is remembered: Inspect memory, correct facts, pin useful entries, retire or restore material, and erase selected records. Review queues help you examine new versions and make deliberate decisions in bulk.

Covenant, Saga extraction, and several retrieval features are opt-in. You choose which to enable; Campaign-scoped recall can narrow which memories are used for a project. Reviewing an entry does not promote an assistant's suggestion into operator-confirmed context.

Find the detail. See the larger picture.

Bring relevant material into the conversation.

Retrieval-Augmented Generation (RAG) gives the model relevant passages from your own material before it answers. Arcanum combines semantic search, exact text search, and layered summaries for different kinds of questions.

Search by meaning · The Weave and Divination: The Weave stores embeddings: numerical representations of text. Divination uses them to find relevant workspace passages and session attachments, even when the wording differs from your question.
Explore larger collections · The Tapestry: RAPTOR-style hierarchical retrieval builds layers of summaries over indexed workspaces, attachments, and session history. This helps connect detailed passages with the themes of a larger body of material.
Find exact words · SQLite FTS5: Local full-text indexes help find specific names, phrases, and records quickly. They complement semantic retrieval when you know the text you are looking for.
Keep the source within reach: Retrieved passages and summaries remain derived context. Finding a relevant match does not make it a new instruction or give it authority beyond its source.

Semantic retrieval needs a configured embedding provider. Workspace retrieval, attachment retrieval, and The Tapestry have their own feature settings, so you decide which material to make available.

OATH · The rule behind the memory

Memory cannot outrank its origin.

OATH means Origin-Bound, Authority-Conserving Transactional History. It governs how memory is kept, found, used, and changed.

If you confirm a project fact, the assistant guesses it, and a document mentions it, those are three different kinds of evidence. OATH keeps that difference visible, even after the text has been summarized or recalled many times.

Know where it came from: A memory and its derivatives retain links to their source, revision, and scope.
Keep its authority in place: Retrieval, repetition, and summarization cannot promote a suggestion into confirmed context or widen where a memory may be used.
Publish related changes together: An assistant result and its staged memory changes commit together, so a failed publication cannot leave only half of that record.
Make changes traceable: Revisions and retirement records preserve the history needed to understand how remembered claims changed.

A search match is a candidate for context, never permission to act.

Before protected context goes to a provider or external tool, Arcanum records what was authorized for that destination. That record helps account for disclosure; erasing a local memory cannot recall a copy already held elsewhere.

Source tracking, separated authority, transactional publication, and disclosure records are implemented. OATH continues to develop; the newer memory capabilities below have not yet reached the main branch.

Memory · In development

Help memory stay useful as the work grows.

These features are implemented on the memory-enhancement development branch and are not yet part of main. Campaign Rollups and Long Rest require explicit enablement.

Carry project context forward · Campaign Rollups: Build a concise, revisioned summary across a Campaign's conversations. Source and sensitivity information travel with it so the summary keeps the limits of the material it came from.
Consolidate observations · Long Rest: Identify duplicate or superseded Saga observations and adjust which ones are recalled. Keep their source history and records explaining each change.
Keep recall relevant: Rank Saga memories using measured usefulness, recency, and clear eligibility rules. Pinned memories and facts you state directly are protected from being set aside just because they are old.
Check whether memory helps · Proving Grounds: Compare the same task with and without selected memory versions. Reports can show a benefit, harm, no measurable difference, or insufficient evidence. Saved reports omit prompts, memory text, and answers.

From a request to a lasting record

See how the pieces work together.

Check the request
Choose a provider
Prepare allowed context
Run models and tools
Commit result and memory
Keep the record

Arcanum checks the request and its limits, prepares relevant context, runs the work, and commits the result with any staged memory changes.

Fully Native AOT compatible

A native engine with less to install.

Arcanum's Native AOT builds start quickly and run without a separately installed .NET runtime. The supported build targets are Apple Silicon macOS, Windows x64, and Windows arm64.

Local records. Deliberate connections.

Keep control of your data and your tools.

The SQLCipher-encrypted Grimoire stores conversations and durable records on your machine. Attachments, uploads, and batch artifacts have separate encryption. Arcanum listens on loopback by default and protects its working API with authentication.

You choose the providers and integrations that receive context. For Campaign work, enabling Sanctum lets you restrict tools, filesystem paths, and network targets. Wards record server-executed tool activity for inspection.

Set spending limits and inspect recorded usage, manage backups and retention, and rotate encryption keys. Provider reporting determines which usage and costs can be measured; subscription CLI usage may be unreported.

Compendium · Configure Arcanum visually

Spend less time editing settings by hand.

Compendium is the desktop editor for arcanum.json. Set up providers and models, choose memory and retrieval features, and adjust security, schedules, pricing, and retention from a visual interface.

It works directly with configuration, so you can prepare your setup before starting the Arcanum service.