jsingletonai/dejavu

Deja Vu is a memory layer for AI tools that runs on your own computer. The pitch in the README is that most AI assistants forget everything between sessions, and the ones that do remember store your context on their servers. This project sits in the middle: you keep your notes, preferences, and prior context in a local SQLite file, and any tool that can call Python, REST, the command line, or MCP can read and write to that same store. The practical effect is that you can add a memory from the command line, ask about it from Claude Desktop, and query it from a Python script, all hitting the same database with no cloud sync and no account. The README gives a four-line quick start: pip install dejavu-memory, dejavu init, dejavu add followed by a quoted sentence, and dejavu search with a question. Memories are written to ~/.dejavu and become searchable immediately. One external service is involved: Venice, a privacy-focused LLM API. The engine sends raw text to Venice to extract durable facts and preferences, then writes them to SQLite along with embeddings. When you search, the query is embedded, the closest matches are pulled, and Venice can re-rank them for relevance. The README says you need to set VENICE_API_KEY before running, and notes that aside from those calls, nothing else leaves your machine. There are four ways to use it. A Python SDK for embedding memory inside agents or scripts, a CLI for quick adds and searches, a local REST API bound to 127.0.0.1 that runs with dejavu serve, and an MCP server. The MCP setup is a small JSON snippet you drop into Claude Desktop, Cursor, or another MCP-compatible client config; once added, the agent shares the same memory store as your CLI and Python code. The project layout is small: a dejavu folder with the core SDK and engine, a cli folder, docs, examples, and tests. Telemetry is off by default and logs are opt-in. The README credits the open-source mem0 project as the basis and the license is Apache 2.0.