Memento

Why we included this project

Most agent teams treat improvement as a choice between renting a bigger model and paying for repeated fine-tuning runs. Memento sidesteps that by treating every finished task as an experience worth keeping: the agent gets sharper over time purely by recalling what worked before, with no weight updates. A meta-planner breaks a query into subtasks, pulls the most relevant stored cases, and hands each piece to an executor that runs MCP tools and writes the result back into the case bank. The memory layer can be a plain parameter-free store or a small trained retriever, and the project ships working MCP tooling for search, crawling, document processing, and running code, so it drops into an existing stack rather than forcing a rebuild. If you run task-oriented agents on closed or local models and want continuous improvement without retraining costs, this is a concrete, runnable reference worth a look.