5 min read·Updated 2026-04-13

Glossary

Canonical one-line definitions for the vocabulary used across the docs — bot, skill, plugin, tier, memory bucket, task, and more.

Core

Bot: The GolemCore Bot process — one container, one dashboard, one workspace volume. When the docs say "the bot", they mean this process.
Workspace: The mounted volume at STORAGE_PATH that holds sessions, preferences, skills, memory, traces, and artifacts. Lose the volume, lose all bot state.
Sandbox: The separate volume at TOOLS_WORKSPACE where shell and filesystem tools operate. Kept distinct from the workspace so tool output cannot corrupt bot state.
Session: One continuous conversation between a user and the bot, persisted under workspace/sessions/ with its full tool-call history.
Turn: One request/response cycle inside a session: input in, tool loop, response out.
Channel: Any entry point that can start a turn — dashboard chat, Telegram, webhooks, Hive command flows.

Terms are grouped by area. Each entry is one sentence — follow the linked concept page when you need the full picture.

Skill: A sticky overlay on an active session: instructions plus optional MCP server and variables. Shapes how the bot behaves without adding new capabilities.
Skill variables: Named values a skill can reference with ${VAR}. Resolved at startup from the skill's vars.json, then the workspace-wide variables.json, then the OS environment.
Plugin: An installable pack that adds a new capability — browser, search, mail, voice, RAG. Enabled, updated, or removed from Settings → Plugin Marketplace.
MCP server: An external tool server launched from a skill under the Model Context Protocol. Its tools are exposed alongside built-in tools while the skill is active.
Tool: A function the bot can call during a tool loop. May come from the core bot, a plugin, or an MCP server.
Tool loop: The inner cycle inside a turn where the model repeatedly requests tools, the bot executes them, and the results are fed back until the model stops asking.

Tier: An abstract model role — balanced, smart, coding, deep — resolved to a concrete provider and model through the model router.
Model router: The mapping from tier to concrete model, stored in preferences/model-router.json. Changed without touching the skills or prompts that reference tiers.
Provider: An upstream LLM API (Anthropic, OpenAI, Groq, a local server). Configured once in preferences/llm-providers.json; referenced from the router by name.
Dynamic escalation: Automatic tier bump when the current model cannot make progress (context too long, repeated tool failure). Always upward, never silent — visible in Sessions.

Memory note: A single fact, preference, or how-to the bot has written down. Stored in one of the four buckets and retrieved when relevant.
Memory buckets: The four places a note can live: working (current task), episodic (dated events), semantic (stable facts), procedural (how-to recipes).
Memory preset: A named bundle of memory settings picked in Settings → Memory. Seven ship with the bot — coding_balanced is the default.
Token budget: The cap on how much memory the bot loads into the prompt at the start of a turn. Set per preset as a soft target and a hard ceiling.
Promotion: Moving a note from a short-term bucket (working, episodic) into long-term memory (semantic, procedural) when it has proven durable. Per-preset toggle.
Decay: Automatic drop of notes that have not been touched within the decay window. Keeps the notebook from bloating with stale facts. Per-preset toggle.
Trace: A detailed, replayable snapshot of a turn — tool calls, inputs, outputs, timings. Inspected from Sessions.
Diary: Auto Mode's append-only log of observations from each goal iteration. Carries context across scheduled runs without reloading everything.

Auto Mode: Scheduled, goal-driven execution loop. A goal owns a task list and a diary; a cron trigger advances the next task.
Goal: A long-running objective in Auto Mode with its own task list and diary. Exists independently of any one session.
Task: A unit of work under a goal. Statuses: PENDING, IN_PROGRESS, COMPLETED, FAILED, SKIPPED.
Delayed action: A way for the bot to schedule future work in its own session — REMIND_LATER, RUN_LATER, NOTIFY_JOB_READY. The bot requests the timer; the runtime fires it.
Custom hook: A named webhook endpoint that turns an incoming HTTP payload into a bot prompt. The message template uses {json.path} placeholders to pull values from the body — not Mustache.

Hive: Optional fleet orchestrator that coordinates many bots through approvals, lifecycle signals, and shared inspection. Not required for single-bot deployments.
Dashboard: The web UI served by the bot on port 8080. The main control surface for chat, settings, and the marketplace.
Preferences: The stack of JSON files under workspace/preferences/ that hold bot configuration. The dashboard is the normal editor; the files are the source of truth.

Overview

The shape of the bot and where each primitive lives.

Troubleshooting

Symptom-to-fix lookup for operational problems.

Overview

What GolemCore is, how the runtime is shaped, and where to go next in the documentation.

User Guide

Quickstart

Get GolemCore running on your machine in five minutes: one Docker command, sign in, add a provider key, send a message.