Endless Galaxy Studios

Context fades.#

Memory doesn't.#

The Problem? Every Session Starts From Scratch#

AI coding agents like Claude Code, Cursor, and Windsurf suffer from a fundamental limitation: context doesn't survive between sessions. Every time you start a new conversation, your coding agent greps and globs files to rebuild what it already knew yesterday — wasting precious tokens in the process.

  • Re-reads the codebase every time — rebuilding context it already had
  • Lost decisions — you remember why you chose that pattern, what you tried and rejected, but your agent doesn't
  • No code awareness — sees files as text, not functions and modules with relationships between them

The Solution: Watches, Learns, Remembers#

Neuroloom hooks into your coding sessions and observes architectural changes, decisions made, and patterns followed. You don't tag or log anything — it watches the work as it happens and figures out what matters.

Watch#

Neuroloom observes your coding sessions in real-time, capturing architectural changes, decisions, and patterns as they happen. No manual tagging required.

Learn#

Observations are distilled into typed memories connected through real relationships. CodeWeaver maps your codebase to the function level and roots each memory in the code it touches — building a knowledge graph over time.

Remember#

Next session, your coding agent starts with the knowledge graph — what was decided, why, and where it lives in your codebase. No re-crawling; no stale CLAUDE.md to maintain. Your agent picks up where it left off.

Key Features#

  • Knowledge Graph — Memories and relationships link to each other. Decisions connect to patterns, patterns connect to code. Your agent sees how everything relates.
  • CodeWeaver — Maps your codebase down to the function level. Roots every memory in the code it touches. Your agent knows where things live, not just what was decided.
  • Hybrid Search — Combines semantic similarity (60%), keyword matching (30%), and importance scoring (10%) for precise retrieval in milliseconds.
  • 17 Memory Types — Decisions, patterns, bug fixes, architecture, integrations, performance optimizations, and more. Each type drives context injection for relevant work.
  • 9 Relationship Types — References, caused_by, supersedes, contradicts, validates, and more. The graph is navigable from any starting point.

Integration#

Built for Claude Code with a first-party plugin that handles session lifecycle, context injection, and memory extraction automatically. Also MCP-compatible with Cursor, Windsurf, and other coding agents via direct .mcp.json configuration.

CodeWeaver runs locally — your source code never leaves your machine. Only extracted memories and relationship metadata are stored in the cloud.

Technical Architecture#

The backend is built on FastAPI with async SQLAlchemy, PostgreSQL 17 with pgvector for vector similarity search, and Redis for caching and background task queuing. The frontend uses React 19, Vite, TailwindCSS 4, and Cytoscape.js for interactive graph visualization.

The MCP server integrates directly with Claude Code, Cursor, and Windsurf — capturing observations automatically and injecting relevant memories when sessions start. A REST API provides programmatic access for custom integrations and CI pipelines.