privacy by architecture

Your data is stored locally. You choose what leaves.

Retrace is built on a simple principle: your most personal conversations are structured on your own machine, and you decide exactly what each tool can see.

Your data is stored locally

SQLite database in ~/.retrace/. No accounts, no telemetry. Conversations, topics, embeddings, and reflections all live on your laptop.

~/.retrace/retrace.db One file. Your file. Copy it, back it up, delete it.

You choose the intelligence

Reflection is optional and never tied to one provider. Connect Claude, Codex, or a model you run yourself over MCP. Retrace works on its own first, the AI is something you bring.

MCP · Claude · Codex · local models

Open source, forever

MIT licensed. Read the code, audit the extension, fork it, build on it. Your reflection infrastructure should never be a black box.

MIT licensed · Open source · Local first

The complete data path

sources

ChatGPT, Claude, WhatsApp, Messenger, Instagram, SMS, Obsidian, Calendar, Keep.

import

Extension, drag-and-drop, live sync. All local. No data sent anywhere.

process

On-device LLM extracts topics, summaries, embeddings. Runs on your hardware.

store

One SQLite file on your laptop. That's it.

When data leaves your machine

Most of Retrace works entirely offline: search, network graph, resources, topic extraction, timeline browsing, Life Rivers, Time Capsules. Data only travels when you point an AI app at it over MCP, and that app can be a model you run locally. When you do use a hosted one, it happens under that provider's policy, and you control exactly what it can access.

Anthropic (Claude)
Default for reflection

Used for embedded chat and MCP server, via your existing Claude subscription. Anthropic does not train on your data. Conversations are not stored beyond the session.

Anthropic privacy policy →
Google (Gemini)
Optional for automations

Used for scheduled reflections (weekly journal, project pulse). Free tier via AI Studio API key. Google does not use API data to train models.

Gemini API terms →
No provider
Fully offline

Without a subscription or API key, reflection features are unavailable. Everything else works: browse, search, network graph, topics, Life Rivers, Time Capsules, labels, Obsidian sync.

You control what providers see

Retrace has a label-based permission system. Tag conversations with labels (platform, topic, or custom), then set allow and block rules per consumer. Claude sees only what you allow.

Embedded chat block private family

Claude cannot access conversations tagged "private" or "family."

MCP server allow chatgpt claude

Claude Desktop can only access your AI conversations, nothing else.

Technical details

Database ~/.retrace/retrace.db
Raw payloads ~/.retrace/raw/
LLM models ~/.retrace/models/
Telemetry None. Zero. Nothing phones home.
Analytics None. We don't know you exist.
License MIT