A companion grown through interaction.
Grounded in your reality.

Most AI companions simulate personality by stuffing hundreds of adjectives into a prompt ('You are an affectionate, playful tsundere...'). This invariably collapses after a few dialogue turns into generic assistant flattery. Siduri starts with an empty relational slate. She has no pre-cooked backstory. Her voice, humor, and boundaries evolve organically through accumulated interaction and verified memories.

When interacting with you, Siduri stages extracted statements as PENDING candidates. These candidates are strictly quarantined from active cognition. Only when you—acting as the human Anchor—explicitly approve a claim does it enter permanent local state in siduri.sqlite (routed directly to Self directives or sovereign Knowledge). Malicious prompt injections attempting to rewrite her beliefs are halted at the gate, unable to corrupt her worldview.

Subjective interaction history and relational directives (@sidurijs/self and @sidurijs/archive) represent the companion's relational bond and episodic records. However, factual user records—such as game rosters, expenses, calendars, and file paths—are objective data. Asking language models to do arithmetic across fuzzy vector embeddings often leads to calculation errors. The sovereign Life Database (@sidurijs/knowledge) provides structured SQLite tables that can be queried deterministically, avoiding the arithmetic drift of probabilistic models.

Yes. Siduri's cognition is decoupled from any single model provider. Her identity baseline, relational stances, and approved directives are compiled into concise, model-neutral inference frames by @sidurijs/self. You can switch between Claude, Gemini, DeepSeek, or local models via Ollama at any time—her memories and bond remain completely intact in your local siduri.sqlite database.

Character cards collapse seven distinct cognitive concepts (identity, tone, directives, self-knowledge, user facts, preferences, and world lore) into an unranked prompt document. This leads to authority inversion and persona drift. Siduri's .self architecture compiles only the relevant slice at inference time: the immutable anchor, current relational stance, channel-specific active directives, and concrete dialogue exemplars.

In @sidurijs/hands, tools are executed through cryptographically signed capability tokens with two-phase reservation. Destructive or external actions require durable human operator authorization. Even if the process or machine restarts, existing approvals are preserved without permitting duplicate execution.

Every organ in Siduri implements the canonical @sidurijs/core runtime protocol and publishes an organ-manifest.json. Developers can create custom organs with standard Node.js ESM exports. In addition, @sidurijs/hands connects natively to Model Context Protocol (MCP) servers, allowing your companion to control local tools and APIs safely.

Siduri is currently in an active, experimental development phase with breaking changes expected. While organ unit tests, packaging, and invariant test suites pass, the framework is not yet tested at full real-world end-to-end (E2E) scale across heterogeneous hardware and long-term usage. APIs, configurations, and database migrations will continue to evolve. We warmly welcome early developer experimentation, issues, and PR contributions.

Because building a reliable, non-drifting companion architecture is a foundational research challenge that benefits from open scrutiny. The CLI and repository are open so technical developers and researchers can clone the codebase, inspect the organ manifests, test the contracts locally, critique our RFCs, and help build a healthier, sovereign companion ecosystem together.