Glossary
Agency
Section titled “Agency”Proactive mode where your agent watches file changes and surfaces relevant vault patterns without being asked. Enable with agency_enable, configure watch paths and thresholds. See Capabilities — Agency.
Anti-pattern
Section titled “Anti-pattern”A known bad approach stored in the vault. Surfaces in searches to warn you away from repeating past mistakes. See types under Entry Types.
Auth level
Section titled “Auth level”Permission required to call an operation — read (query data), write (add/modify data), or admin (delete data, reset state). Enforced at tool registration. See Security & Privacy.
The intelligence layer on top of the vault. Tracks pattern strength, maintains TF-IDF vocabulary, and provides recommendations for plans. See Under the Hood.
Chat facade
Section titled “Chat facade”The facade for chat transport integration — sessions, authentication, response chunking, voice transcription/synthesis, browser isolation, notifications, and message queue. Powers Telegram bots, web chat, and other conversational interfaces. See Capabilities — Chat.
Cognee
Section titled “Cognee”Optional open-source knowledge graph engine that adds vector similarity search alongside TF-IDF. See Cognee Integration.
Cognify
Section titled “Cognify”The process where Cognee converts vault entries into vector embeddings and knowledge graph connections. Runs automatically with debounced scheduling after captures.
Compound loop
Section titled “Compound loop”The self-reinforcing cycle: capture → vault → brain → plans → work → knowledge extraction → vault. Each cycle makes the next one better. See The Knowledge-Driven Approach.
Context facade
Section titled “Context facade”Entity extraction, knowledge retrieval, and context analysis. Analyzes prompts to extract files, functions, domains, and technologies, then retrieves relevant knowledge from vault, Cognee, and brain. Used internally by the orchestrator. See Capabilities — Context.
Curator
Section titled “Curator”Automated maintenance system that keeps the vault clean — deduplication, contradiction detection, decay scanning, health audits, tag normalization, LLM enrichment. See Under the Hood.
The gradual reduction of an unused pattern’s strength score over time. Ensures stale knowledge doesn’t outrank actively useful patterns.
Domain
Section titled “Domain”A knowledge area (e.g., frontend, security, infrastructure). Each domain gets its own search partition and facade with 5 operations. Added via npx @soleri/cli add-domain <name>. See Customizing Your Agent.
Domain facade
Section titled “Domain facade”An MCP tool entry point scoped to a single domain. Each domain facade exposes 5 operations: get_patterns, search, get_entry, capture, remove.
The difference between what a plan intended and what actually happened during execution. Measured during reconciliation. Low drift means the plan was accurate; high drift means reality diverged. See Planning.
Entry types
Section titled “Entry types”The kind of knowledge stored in the vault: pattern, anti-pattern, rule, playbook, workflow, principle, reference.
Facade
Section titled “Facade”A single MCP tool entry point that dispatches to multiple operations via the op parameter. Every agent has 13 semantic facades plus one per domain. See API Reference.
SQLite’s Full-Text Search extension, version 5. Powers the vault’s text search with porter tokenizer for stemming.
An approval checkpoint in a plan or loop. Plans have a two-gate system (approve plan, then approve tasks). Loops use gates to decide whether to continue iterating.
Governance
Section titled “Governance”Controls how knowledge enters the vault — capture quotas, proposal gates, duplicate detection, and decay policies. Presets: strict, moderate, permissive. See Customizing Your Agent.
A quality gate that runs automatically during development. Checks code against rules (e.g., no console.log, no any types) and blocks violations. Installed via npx @soleri/cli hooks. See Customizing Your Agent.
Intake
Section titled “Intake”The process of ingesting external content (PDFs, books) into the vault. Content is chunked, analyzed, and converted into structured knowledge entries. See intake_ingest_book in Capabilities — Vault.
Knowledge extraction
Section titled “Knowledge extraction”The automatic process where the brain examines a completed plan session — tools used, files modified, outcomes — and proposes reusable patterns back into the vault.
Knowledge pack
Section titled “Knowledge pack”A bundle of pre-built knowledge entries (patterns, anti-patterns, principles) that can be installed into an agent. See Customizing Your Agent.
An iterative validation cycle (do → validate → fix → repeat) that converges on a quality target. Built-in modes: component-build, plan-iteration, custom. See Validation Loops.
MCP (Model Context Protocol)
Section titled “MCP (Model Context Protocol)”The protocol that connects your agent to Claude Code. Your agent registers tools over MCP; Claude Code calls them based on your conversation.
MCP Bridge
Section titled “MCP Bridge”A chat facade subsystem that allows local tool execution from chat platforms (Telegram, web). Registered tools can be called via the bridge with output compression for display. See Capabilities — Chat.
Pattern
Section titled “Pattern”A proven approach stored in the vault. Surfaces in searches and brain recommendations to guide future work.
Pattern strength
Section titled “Pattern strength”A score (0.0–1.0) tracked by the brain. Increases when a pattern is used successfully; decreases with disuse (decay) or drift. Higher strength means more confident recommendations.
Playbook
Section titled “Playbook”A multi-step procedure with validation criteria at each step. Created from repeated workflows, run on demand. Supports start, step, complete, and match operations. See Planning.
RateLimiter
Section titled “RateLimiter”Transport-level request throttling. Tracks per-client request counts within configurable time windows and blocks requests that exceed the limit. See Transports.
Reconciliation
Section titled “Reconciliation”The step after plan execution where the agent compares what was planned vs. what happened. Produces a drift report and accuracy score. Feeds the knowledge loop.
SessionManager
Section titled “SessionManager”Transport-level session tracking. Manages client sessions with TTL-based expiry, unique ID generation, and automatic reaping of expired sessions. See Transports.
Severity
Section titled “Severity”Priority level of a vault entry: critical (must follow), warning (should follow), suggestion (nice to have). Affects search ranking — critical entries always surface first.
TF-IDF
Section titled “TF-IDF”Term Frequency–Inverse Document Frequency. A text ranking algorithm that weights terms by rarity. “Authentication” scores higher than “the” because it’s more meaningful. The brain maintains a TF-IDF index across all vault entries.
Transport
Section titled “Transport”The communication layer between your agent and its clients. Soleri supports four transports: stdio (MCP for Claude Code), HTTP/SSE (REST APIs with Server-Sent Events), WebSocket (bidirectional streaming), and LSP (Language Server Protocol for editors). See Transports.
The agent’s long-term knowledge store. A SQLite database with FTS5 full-text search. Stores structured entries (patterns, anti-patterns, rules) organized by domain and severity. Supports branching, scoping, sharing, and multi-tier connections. See Under the Hood.
Vault branch
Section titled “Vault branch”An isolated copy of vault state for experimentation. Create a branch, add entries, and merge back — or delete without affecting the main vault. See Capabilities — Vault.
Back to Your Agent — Quick Reference or explore Capabilities for the full operation list.