spacedriveapp/spacebot

What it solves

Spacebot is an agent infrastructure designed to move beyond the limitations of single-session AI agents. It solves the problem of "monolithic" agent loops where a single LLM thread handles conversation, thinking, and execution simultaneously, which often leads to responsiveness issues and loss of context in long-running sessions.

How it works

Spacebot uses a specialized process architecture to delegate tasks across five distinct process types:

• Channels: User-facing processes that maintain identity and personality but delegate work to other processes.
• Branches: Concurrent forks of a channel's context used for reasoning and decision-making without blocking the user.
• Branches: Independent processes that execute actual work (e.g., shell commands, browser automation) using the full context of the intent.
• The Compactor: A monitor that manages context size by creating "Chronicles"—append-only checkpoints of session history—rather than lossy rolling summaries.
• The Cortex: A system supervisor that manages workers, enforces timeouts, and maintains the memory graph.

Who it’s for

It is built for teams, communities (e.g., Discord servers), and solo users who need an agent that can handle concurrent multi-user conversations, shared memory across channels, and autonomous background work.

Highlights

• Typed Memory Graph: Uses SQLite and LanceDB to store facts, preferences, and decisions as a typed graph with hybrid vector and full-text search.
• Autonomous Task System: Operates via "wakes" (scheduled or event-driven triggers) and a goal-driven model where tasks require user approval before execution.
• Self-Improving Skills: Automatically distills successful multi-step work into reusable "skills" that are injected into future worker prompts.
• Extensive Integration: Native adapters for Discord, Slack, Telegram, Twitch, and Email, plus MCP (Model Context Protocol) support for external tools.
• Security-First Design: Implements OS-level filesystem containment (bubblewrap on Linux, sandbox-exec on macOS) and encrypted secret storage to isolate LLM processes from system credentials.