ricoeur

A local-first archive, search, and intelligence engine for your LLM conversation history.

Named after Paul Ricoeur, whose work on narrative identity argued that we understand ourselves through the stories we construct from our lived experience. ricoeur reconstructs the narrative of your intellectual life from thousands of AI conversations.

██████╗ ██╗ ██████╗ ██████╗ ███████╗██╗   ██╗██████╗
██╔══██╗██║██╔════╝██╔═══██╗██╔════╝██║   ██║██╔══██╗
██████╔╝██║██║     ██║   ██║█████╗  ██║   ██║██████╔╝
██╔══██╗██║██║     ██║   ██║██╔══╝  ██║   ██║██╔══██╗
██║  ██║██║╚██████╗╚██████╔╝███████╗╚██████╔╝██║  ██║
╚═╝  ╚═╝╚═╝ ╚═════╝ ╚═════╝ ╚══════╝ ╚═════╝ ╚═╝  ╚═╝
              ──────────────────────────
                your conversation archive

The wordmark above is painted on as a brief opening animation when you launch ricoeur tui. Press any key to skip it, or set RICOEUR_NO_SPLASH=1 to disable.

Quickstart

# Install with uv
uv sync

# Initialize the database
uv run ricoeur init

# Import your ChatGPT export
uv run ricoeur import chatgpt ~/Downloads/chatgpt-export/conversations.json

# ...or your Claude export (Settings > Privacy > Export data)
uv run ricoeur import claude ~/Downloads/claude-export/conversations.json

# ...or your Claude Code sessions — no export needed, they're already on disk
uv run ricoeur import claude-code

# Build the intelligence layer (language detection, embeddings, analytics)
uv run ricoeur index

# See what you've got
uv run ricoeur stats

# Search your history
uv run ricoeur search "thermal simulation"

Commands

Search

ricoeur supports three search modes:

When embeddings are available (after ricoeur index), search automatically uses hybrid mode. If no embeddings exist, it falls back to keyword search. Flags like --code or --role also force keyword mode since they rely on FTS5.

# Hybrid search (default — combines keyword + semantic)
ricoeur search "deployment strategies"

# Force keyword-only (FTS5)
ricoeur search "streamlit dashboard" --keyword

# Force semantic-only (cosine similarity)
ricoeur search "how to containerize apps" --semantic

# Filter by platform, language, date
ricoeur search "error fix" --platform chatgpt --lang en
ricoeur search "strategie marketing" --lang fr --since 2025-01-01

# Search only in code blocks (auto-uses keyword mode)
ricoeur search "import pandas" --code

# More output formats coming soon: json, full, ids

Why semantic search?

Keyword search only finds exact word matches. Semantic search finds conceptually related conversations — even when the exact words don't appear.

Query: "how to containerize applications"

$ ricoeur search "how to containerize applications" --keyword
Found 3 results for "how to containerize applications" (keyword)

$ ricoeur search "how to containerize applications" --semantic
Found 20 results for "how to containerize applications" (semantic)
 #   Score    Date        Title
 1   0.5515   2025-07-29  Free docker deployment options
 2   0.5152   2026-01-28  Secure Clawdbot Setup
 3   0.5081   2025-03-11  Docker noVNC Setup
 4   0.5014   2025-09-21  HTML upload and serve
 5   0.4782   2022-12-27  Wasm vs Container Comparison
 ...

Keyword found 3 results matching the literal words. Semantic found 20 — including Docker, Wasm, and container deployment conversations that never mention "containerize applications".

Terminal UI

Prefer browsing interactively? Launch the TUI:

# Install the optional dependency
uv sync --extra tui

# Launch
uv run ricoeur tui

It opens to your most recent conversations. Type a query and press Enter to keyword-search; clear the box and press Enter to return to the recent list. Select a row (Enter) to read the full conversation, rendered as Markdown.

Index

After importing, build the intelligence layer:

# Run all layers: language detection, embeddings, analytics
ricoeur index

# Second run skips what's already cached
ricoeur index

# Force a full rebuild
ricoeur index --force

# Run specific layers only
ricoeur index --embeddings
ricoeur index --analytics

# Use a different embedding model
ricoeur index --embed-model ollama:nomic-embed-text
ricoeur index --embed-model st:all-MiniLM-L6-v2 --device cpu

Each layer requires its optional extra:

# Install all index dependencies at once
uv sync --extra langdetect --extra embeddings --extra analytics

Import options

# Re-import safely (updates existing, adds new, never deletes)
ricoeur import chatgpt conversations.json --update

# Dry run — parse and validate without writing
ricoeur import chatgpt conversations.json --dry-run

# Only import recent conversations
ricoeur import chatgpt conversations.json --since 2025-01-01

# The same flags work for Claude exports
ricoeur import claude conversations.json --update

ChatGPT and Claude exports are both supported. Each platform ships a conversations.json (sometimes inside a .zip) — point ricoeur at either the JSON file or the zip and it will find the conversations.

Claude Code sessions

Claude Code needs no export step. It writes every session to ~/.claude/projects/<project>/<session>.jsonl as it works, so ricoeur reads them straight from disk:

# Import every session (defaults to ~/.claude/projects)
ricoeur import claude-code

# Just one repo's sessions
ricoeur import claude-code --project ricoeur

# A single project directory, or one session file
ricoeur import claude-code ~/.claude/projects/-Users-me-Devel-ricoeur

Sessions are append-only and may be mid-write, so re-running the import is cheap and safe: unchanged session files are skipped, a session that has grown gains only its new messages, and nothing is ever replaced or deleted. Run it again whenever you want to catch up.

These transcripts are archived as the platform claude-code, tagged with the repo they ran in:

ricoeur search "migration" --platform claude-code
ricoeur stats --project ricoeur

What gets archived is conversation: your prompts, Claude's replies, and the code it wrote (Bash commands, Write contents, Edit diffs — all searchable via --code). Tool output fed back to the model is dropped by default, since it would outnumber the actual conversation roughly 30 to 1 and turn every --role user search into a wall of command output. Two flags opt back in:

# Keep tool output too — a complete audit trail, much larger
ricoeur import claude-code --include-tool-results

# Include subagent (Task/Agent) transcripts
ricoeur import claude-code --include-sidechains

Coming soon: Gemini and custom JSON imports.

Optional extras

Install additional capabilities as needed:

# Language detection
uv sync --extra langdetect

# Semantic search with sentence-transformers
uv sync --extra embeddings

# Topic modeling with BERTopic (coming soon)
uv sync --extra topics

# Analytics with DuckDB + Parquet
uv sync --extra analytics

# Terminal UI
uv sync --extra tui

# MCP server for Claude Desktop (coming soon)
uv sync --extra mcp

# Web API server (coming soon)
uv sync --extra serve

# Everything
uv sync --extra all

Configuration

Config lives at ~/.ricoeur/config.toml:

[general]
home = "~/.ricoeur"
default_language = "en"

[embeddings]
model = "st:paraphrase-multilingual-mpnet-base-v2"
batch_size = 64
device = "auto"

# Topics and summarize config — coming soon
# [topics]
# min_cluster_size = 15
# n_topics = "auto"
#
# [summarize]
# enabled = false
# model = "ollama:llama3.2"

Override the data directory with the RICOEUR_HOME environment variable.

Architecture

~/.ricoeur/
├── config.toml          # Configuration
├── ricoeur.db           # SQLite database (FTS5 search)
├── analytics/           # Parquet files for DuckDB
├── embeddings/          # Sentence-transformer vectors
├── models/              # Saved BERTopic models (coming soon)
└── attachments/         # Extracted files (coming soon)

License

MIT