Remembrane lets you run agent memory from a single SQLite file

Most agent memory solutions are overkill. If you just need an LLM to remember a few thousand facts across sessions, setting up a hosted API or a full-blown vector database feels like using a sledgehammer to crack a nut. I've been looking for something lightweight, and Remembrane hits that spot by keeping everything in one SQLite file with zero default dependencies.

The biggest win here is that recall is deterministic. In a real-world AI workflow, being able to write unit tests that assert exactly what an agent remembers—and running those in CI—is huge. Most memory layers are black boxes, but this allows for actual verification. Plus, since it's just a file, you can copy it, inspect it, or wipe it without managing a server process.

How it handles retrieval

The ranking system isn't just a simple similarity search. It uses a weighted formula:

• Similarity: The base match.
• Recency: Newer memories can be prioritized (this is configurable).
• Importance: Higher weight for critical facts.
• Utility: Boosts memories that have been useful in previous turns.

It also includes a heuristic check to flag contradictory memories. It's not perfect "ground truth," but it's a great way to find candidates for review.

Technical Trade-offs and Benchmarks

If you're planning a deep dive into this, you need to know the constraints. The default embedder is lexical. If you need true semantic recall, you have to plug in OpenAI or sentence-transformers. Also, this is designed for agent-scale memory (thousands of items). Once you hit around 50k entries, you've outgrown this and should move to a dedicated vector DB.

I did a quick benchmark against mem0 using no-LLM mode (infer=False) to compare the storage and ranking layers specifically. Using the same embedder, Remembrane was faster, took up less disk space, and was more accurate at returning updated facts because of the recency weighting.

Integration and Deployment

For those looking for a practical tutorial on getting this into a project, it's pretty straightforward. It exposes an MCP server, meaning Claude can use it directly. There are also adapters for LangChain and CrewAI.

If you want to implement it from scratch in a python environment, the setup is minimal:

pip install remembrane

Then you can initialize the store and start adding memories. Because it's SQLite-backed, the deployment is as simple as moving a .db file.

The current gaps are mainly around diversity-aware re-ranking (to avoid near-duplicate results) and the CrewAI adapter, which is currently a helper rather than a full backend. For anyone building a custom LLM agent who is tired of infrastructure bloat, this is a very pragmatic approach to persistence.