Dongtian

中文文档

A local, unified memory layer for AI conversations — stored in one SQLite file and exposed as an MCP server.

Dongtian helps you keep and search conversation history across multiple AI tools (e.g. Claude Code, Codex, ChatGPT exports, OpenCode, Slack).

It ingests conversations into a simple SQLite schema and provides:

• keyword search (SQLite FTS5, offline)
• optional embedding search (via an embedding API)
• hybrid ranking (keyword + embedding)
• optional remote sync over SSH
• an optional “cave survey” knowledge graph (deposits + passages)

---

What it stores (terminology)

Dongtian uses a cave metaphor to keep the data model small and consistent:

• Cavern: the SQLite database file
• Layer: a top-level grouping (often a machine name or source, e.g. codex-laptop)
• Chamber: a session/topic inside a layer (often a date or session id)
• Stratum: a chunk of text inside a chamber (ingested content)

Optional cave survey (knowledge graph):

• Deposit: an entity (person, project, concept, tool)
• Passage: a relationship between deposits (e.g. uses, depends_on)

---

Quick Start

Install

pip install dongtian

Or from source (recommended while iterating):

git clone https://github.com/siaochuan/dongtian.git
cd dongtian
pip install -e .

Configure (optional)

Create ~/.dongtian/config.json:

{
  "db_path": "~/.dongtian/cavern.db",
  "embedding_api_key": "YOUR_KEY",
  "embedding_base_url": "https://api.siliconflow.cn/v1",
  "embedding_model": "BAAI/bge-m3",
  "remote_hosts": [
    {"host": "dev-machine", "layer": "remote-dev"}
  ]
}

• Embeddings are optional. If not configured, Dongtian uses keyword search (FTS5).
• For remote_hosts, layer and wing are both accepted as the layer name override.

Use as MCP Server

Claude Code — add to .mcp.json or ~/.claude/settings.json:

{
  "mcpServers": {
    "dongtian": {
      "command": "python",
      "args": ["-m", "dongtian"]
    }
  }
}

Codex CLI — register via command:

codex mcp add dongtian -- python -m dongtian

---

MCP Tools

Tool	Purpose
list_layers	List layers
list_chambers	List chambers in a layer
browse_chamber	Browse strata (paged)
search	Keyword / embedding / hybrid search
ingest_source	Ingest a single file or directory (source-specific)
ingest_claude_project	Bulk ingest Claude Code sessions
ingest_codex_sessions	Bulk ingest Codex sessions
ingest_opencode	Ingest OpenCode (DeepSeek) database
sync_remote	Sync from one remote host via SSH
sync_all_remotes	Sync all configured remote hosts
discover_remote	Discover remote availability (no pull)
add_deposit	Add/get a deposit (entity)
add_passage	Add a passage (relationship)
survey	Query passages (knowledge graph)
extract_survey	Extract deposits/passages from a stratum

---

Supported Sources

Source	Format	What it parses
Claude Code	JSONL	~/.claude/projects/ session histories
Codex	JSONL	~/.codex/sessions/ rollouts (turn-based parsing with tool summaries)
OpenCode (DeepSeek)	SQLite	~/.local/share/opencode/opencode.db
ChatGPT	JSON	OpenAI conversation export (conversations.json)
Slack	JSON	Channel export (directory or single file)
Plain text	.txt / .md	Split on headings or paragraphs

All sources are chunked into conversation turns and then into strata.

---

License

MIT