mem is a small, open-source “memory layer” you can put next to an agent so it can persist context across sessions and projects. Instead of stuffing everything into a prompt, you write memories (notes, facts, decisions, artifacts) into a store that supports similarity search and structured metadata, then retrieve only what’s relevant for a given task.
The project ships as a FastAPI service with a Chroma-backed vector index, plus optional multimodal indexing (e.g., generating image captions so screenshots can be searched like text). The README includes a simple Streamlit UI and a LangChain integration example. If you want to try it quickly, the fastest path is to run the server, add a handful of “project memories” (requirements, constraints, links), and then wire up your agent so every session starts with a retrieval step.