ChatCrystal

The problem: where does experience go after a conversation?

AI coding histories contain more than code. They also hold debugging context, failed attempts and the reasoning behind a decision. When that material stays scattered across tools, the next similar problem begins with another search for context.

ChatCrystal organizes a workflow around turning those conversations into reusable experience: import the source material, distill notes, retrieve related knowledge and bring it into the next task.

Four stages of a knowledge workflow

This is a conceptual workflow, not a promise that every stage runs automatically after import. In particular, remote-instance imports and summarization are separate actions, allowing users to decide when to process content with a model.

Import the context

The project supports local histories from tools including Claude Code, Cursor, Codex CLI, Trae and GitHub Copilot. A shared workspace makes it easier to find related conversations without remembering which client originally held them.

Distill a useful reading layer

Notes include titles, summaries, conclusions, code snippets and tags. The source conversation carries the context; a structured note offers a shorter entry point for deciding whether an experience is relevant.

Retrieve by meaning and explore relationships

Semantic search helps find relevant experience even when the wording differs. The tag graph provides an additional way to explore related topics. Connections based on tag co-occurrence describe associations, rather than proving causation.

Reuse knowledge in the next task

CLI and MCP interfaces let agents recall existing experience and write new records back. Knowledge can participate in another coding workflow instead of remaining a document that must be found and copied manually.

Shared responsibilities, multiple interfaces

The TypeScript repository includes a React client, Fastify server and Electron desktop entry point. Desktop, CLI / Web and Docker self-hosting offer different access patterns. The diagram summarizes responsibilities, not an exact deployment topology.

Choice	Benefit	Tradeoff
Local-first workflow	Organize your own histories and knowledge	Does not imply offline inference
Separate LLM and embedding settings	Choose generation and retrieval providers independently	The model types are not interchangeable
Conversations and notes	Preserve context alongside concise results	Generated conclusions still need judgment
CLI / MCP access	Reuse knowledge in automation and agent tasks	Results must be evaluated for the current task

Content processed by an external LLM or embedding service is sent to that configured provider. The deployment and model choices determine where processing happens.

Start with a familiar set of conversations

Import one project, inspect the source records, generate a small batch of notes and search for a problem you remember solving. This makes it easier to assess the workflow against your own needs.

See the README for installation and configuration, and the development guide for implementation details.

Implementation and workflows

• Five AI conversation data formats: format differences and import limitations.
• Choosing a local vector store: deployment, retrieval, and index-recovery tradeoffs.
• Exporting ChatCrystal notes to Obsidian: a Markdown snapshot workflow that preserves note content and relationships.
• Local-first and cloud AI: data boundaries across storage, inference, and agent clients.
• RAG and knowledge graphs: candidate retrieval, note relationships, and one-hop expansion.
• Knowledge bases for AI agents: recall, validation, and writeback grounded in evidence.