llm-memex 0.12.0

Personal conversation knowledge base: import, search, and analyze conversations from ChatGPT, Claude, Gemini, and Claude Code

llm-memex

Personal conversation knowledge base. Import, search, and analyze conversations from ChatGPT, Claude, Gemini, and Claude Code. MCP-first for LLM agent access.

Install

pip install py-llm-memex

For development:

git clone https://github.com/queelius/llm-memex
cd llm-memex
pip install -e ".[dev]"

Quick Start

Import conversations:

llm-memex import conversations.json          # auto-detects format
llm-memex import ~/.claude/projects/          # directory of Claude Code sessions
llm-memex import export.json --format openai  # force format

Browse and search:

llm-memex show                               # list conversations
llm-memex show <id>                          # view a conversation
llm-memex show --search "topic"              # full-text search

Query the database: use the MCP server (for LLM agents) or sqlite3 directly:

sqlite3 ~/.memex/conversations/conversations.db "SELECT count(*) FROM conversations"

Export:

llm-memex export output.md --format markdown
llm-memex export output.json --format json
llm-memex export ./archive --format arkiv    # universal archive format
llm-memex export ./site --format html        # self-contained HTML SPA

MCP server (for Claude Desktop, agent SDKs, etc.):

llm-memex mcp

Scripts:

llm-memex run --list                                           # available scripts
llm-memex run redact --words "secret" --level word --apply
llm-memex run redact --pattern-file api_keys.txt --apply
llm-memex run enrich_trivial --apply

Supported Formats

Format	Import	Export	Notes
OpenAI (ChatGPT)	Yes	-	JSON export
Anthropic (Claude)	Yes	-	JSON export
Gemini	Yes	-	JSON export
Claude Code	Yes	-	JSONL, conversation-only mode
Claude Code (full)	Yes	-	Full fidelity: tool_use, thinking, subagents
Markdown	-	Yes
JSON	-	Yes
HTML (SPA)	-	Yes	Self-contained, light/dark, librarian chat
Arkiv	-	Yes	Universal record format (JSONL + schema.yaml)

HTML Export

The HTML exporter builds a self-contained single-page app that loads the SQLite database client-side via sql.js (Wasm). Features:

• Light/dark mode: follows OS preference with a manual toggle
• Full browser UI: conversation list, search, filter by source/tag, timeline sparkline
• Librarian chat: ask questions about your archive. An LLM queries the database via Anthropic tool use, using the metafunctor-edge proxy by default (no API key required). You can also configure a direct Anthropic endpoint.
• Per-conversation resume chat: continue an existing conversation
• Marginalia: annotate messages and conversations with free-form notes, inline in the browser

Notes (Marginalia)

Annotate messages and conversations with free-form text notes. Notes are stored in a dedicated notes table with FTS5 search, and appear across all surfaces (CLI, MCP, HTML SPA, exporters).

llm-memex run note add --conv <id> "this was a turning point" --apply
llm-memex run note add --conv <id> --msg <id> "key insight here" --apply
llm-memex run note list --conv <id>
llm-memex run note search "turning point"
llm-memex run note delete <note_id> --apply

Notes are included in exports by default. Use --no-notes to strip them:

llm-memex export ./public --format html              # includes notes
llm-memex export ./public --format html --no-notes   # strips notes

In the HTML SPA, click the pencil icon on any message or conversation header to add a note inline. Notes persist in the browser's sql.js copy and are included when you download the DB.

Multi-Database Config

llm-memex supports multiple named databases via ~/.memex/config.yaml:

primary: conversations
databases:
  conversations:
    path: ~/.memex/conversations
  claude_code_full:
    path: ~/.memex/claude_code_full
  sandbox:
    path: ~/.memex/sandbox

All CLI commands and MCP tools accept --db <name> (CLI) or db=<name> (MCP) to target a specific database. The primary database is used when no name is specified.

MCP Tools

When running as an MCP server, llm-memex exposes 6 tools:

• execute_sql: Primary read interface. All queries via SQL (read-only by default).
• get_conversation: Tree-aware retrieval + export (metadata, messages, markdown/JSON).
• get_conversations: Bulk retrieval with filters (tag, source, model, search, ids) and optional full messages.
• update_conversations: Modify properties, tags, and enrichments (bulk).
• append_message: Add messages to conversation trees.
• add_note: Annotate a message or conversation with a free-form text note.

Resources: llm-memex://schema (DDL + query patterns), llm-memex://databases (multi-db discovery + stats).

Development

pytest tests/llm_memex/ -v             # run tests
pytest tests/llm_memex/ --cov=llm_memex    # with coverage

License

MIT