An external hard drive for your AI's goldfish brain

What it does

DejaVu acts as a persistent memory layer between you and any AI tool that speaks Python, REST, CLI, or MCP. It stores facts and preferences in a local SQLite database under ~/.dejavu, then surfaces them on demand so agents like Claude Desktop or custom scripts can recall context from previous sessions. The only external dependency is Venice’s API, which handles the LLM work for extracting memories and ranking search results.

The interesting bit

The architecture is deliberately boring in the best way: plain SQLite, open file formats, and multiple interfaces hitting the same local engine. The twist is that it uses an external LLM provider not for chat, but for memory hygiene—deciding what is worth keeping and how to rank it—while the actual storage and retrieval stay firmly on your machine.

Key highlights

• Stores memories locally in ~/.dejavu with no hosted account or telemetry.
• Exposes the same SQLite store through Python SDK, CLI, REST API, and MCP server.
• Uses Venice for memory extraction and query re-ranking, but keeps embeddings and data local.
• Designed as a foundation for a portable memory standard rather than a locked-in product.

Caveats

• Requires a Venice API key; LLM calls for extraction and re-ranking leave the machine.
• The README describes this as a “first release” that ships the foundation, so the broader portable-memory standard is still aspirational.

Verdict

Worth a look if you are building agents that need to remember users without building yet another SaaS backend. Skip it if you need a fully offline, air-gapped system—DejaVu still dials Venice for the thinking part.