Canonical definitions for Thane-specific terminology. When documenting or discussing Thane, use these terms consistently.
The universal execution primitive. Every piece of work — conversations, delegations, background watchers — runs as a loop with the same reasoning cycle: context assembly, tag activation, planning, tool execution, response. Three operation modes: request/reply (one-shot, up to ten iterations), background task (detached, result delivered later), and autonomous loop (persistent, self-paced — the agent directs its own attention and controls its own timing). A registry tracks all active loops. See The Agent Loop.
A semantic label (ha, email, forge, web, etc.) that gates which tools
and talents are active in a given session. Sessions start with a minimal set of
always_active tags. Additional tags are activated on demand — by the agent,
the user, or the channel. Tags create delegation pressure: the orchestrator
starts with a small toolset and naturally reaches for delegation when it needs
capabilities outside its active set.
See The Agent Loop.
The protocol Thane uses to sync its contact directory with native apps (macOS Contacts, iOS, Thunderbird). Served on port 8843 via RFC 6352. Contacts managed in Thane are accessible from your phone without export/import. See API & Endpoints.
A context note written by the agent when closing a session. It summarizes key decisions, open threads, and relevant state so the next session can pick up where the previous one left off. The primary mechanism for continuity across session boundaries.
A communication pathway into or out of Thane: the web API, Home Assistant's
Ollama integration, email (IMAP/SMTP), Signal messaging, or MQTT. Each
channel can pin capability tags — an email-originated request automatically
activates the email tag.
A full state snapshot of conversations and metadata, persisted to SQLite. Triggered by message count threshold, graceful shutdown, or manual request. Checkpoints enable crash recovery — the agent resumes exactly where it left off. See Memory.
The process of summarizing older conversation messages to reduce token count while preserving semantic content. Compaction happens automatically when a conversation approaches context limits. The LLM generates the summary, preserving decisions, facts, and preferences.
One of four distinct sections of the system prompt, each with a specific purpose: Persona (identity), Talents (behavior), Inject Files (knowledge), and Session Context (current state). Mixing concerns across layers degrades agent behavior. See Context Layers.
A local model that executes tool-heavy work on behalf of the orchestrator. Delegates receive precise instructions and a full toolset, execute mechanically, and return structured results. They run at zero API cost. See Delegation.
The two-model pattern at the heart of Thane's cost architecture. The orchestrator (cloud or large local) plans and reasons. The delegate (small local) executes tool calls. The smart model thinks; the cheap model does. See Delegation.
A single enforcement point for all outbound messages (email, eventually Signal and other channels). Will provide rate limiting by trust zone, dedup, and content scanning. The most critical planned structural safety control. See Trust Architecture.
A post-session analysis that extracts key facts from conversations into the semantic knowledge base. Enables long-term learning and pattern recognition across sessions.
Anything that can wake the agent loop: an API request, a Home Assistant WebSocket state change, an MQTT wake subscription, an inbound email or Signal message, or a scheduled task firing. See Event Sources.
A reference document loaded into the system prompt's knowledge layer. Contains factual information — user profiles, infrastructure notes, identity documents — not behavioral directives. See Context Layers.
A persistent global behavioral mode that modifies context across all conversations. Unlike capability tags (which activate tools), lenses shape how the agent behaves. Activated and deactivated via tools.
A standard protocol for extending LLM capabilities via external tool servers. Thane hosts MCP servers as subprocesses and bridges their tools into the agent loop. The primary example is ha-mcp, which provides 90+ Home Assistant tools. See Delegation.
A scheduled self-reflection process where the agent analyzes recent
interactions and updates its personality file (ego.md). Runs on a
quality-floor model. Constrained to observation — reads files and
memory but doesn't execute tools or interact with external systems.
The primary model in a delegation interaction. Plans the approach, writes precise instructions for the delegate, and shapes the final response from the delegate's execution results. Typically a cloud or large local model. See Delegation.
A generic namespaced key-value store with TTL support. Persists lightweight operational state: email polling cursors, feature flags, session preferences. Not for structured domain data.
A markdown file (persona.md) that defines the agent's identity —
voice, personality, values, boundaries. Loaded into the system prompt's
identity layer. The persona shapes how the agent interprets ambiguous
requests and communicates.
See Context Layers.
A parameter that influences model selection for a specific request. Hints include quality floor, speed preference, local-only restriction, and mission type. Hints propagate through delegation. See Routing Profiles.
A named preset that maps to a set of routing hints. Selected by setting
the model name in any Ollama-compatible client (e.g., thane:latest,
thane:premium, thane:ops). Profiles describe intent; the router
finds the best model.
See Routing Profiles.
A piece of learned knowledge stored with optional embeddings for natural language recall. Categorized (user, home, device, routine, preference) and confidence-scored. Facts are long-term memory — they persist across sessions, restarts, and model changes. See Memory.
A bounded conversation with its own message history, working memory, and metadata. Sessions can be closed (with carry-forward), checkpointed (for recovery), or split (forking context). Completed sessions are archived with full-text search.
A markdown file that teaches the agent a specific behavior or skill. Talents are tag-filtered — each declares which capability tags it requires and only loads when those tags are active. Transparent, editable, version-controlled. Natural language as the configuration mechanism. See Context Layers.
A classification assigned to every contact: admin, household,
trusted, known, or unknown (implicit). Trust zones are the
universal router for permissions across the system — gating email send,
compute allocation, notification priority, and proactive behavior.
Validated in Go, not prompts.
See Trust Architecture.
A per-session read/write scratchpad for transient context — emotional texture, conversational arc, things worth tracking within a conversation but not permanently. Survives compaction within a session but doesn't persist across sessions. See Memory.