What is mnemo-cognee-hack?

Mnemo is a memory system for AI agents, built as a hackathon project on top of a service called Cognee. The problem it addresses is that most AI agent runs are stateless: each time an agent runs, it has no memory of what happened before, so it can repeat the same mistakes over and over with no way to tell if it is actually improving. Mnemo gives agents a lasting, graph structured memory of how they have behaved in the past, then uses that history to evaluate new runs. Every time an agent completes a task, Mnemo records the task, the tools the agent called, and the outcome into a knowledge graph. When a new run comes in, or when a real incident is reported through Jira, Mnemo looks up similar past runs from that graph and uses them as evidence for an AI judge that decides whether the new run passed or failed, and must point to the specific past runs it used to make that decision. If a person disagrees with a verdict and corrects it, that correction is remembered too, so the next similar run gets judged differently without anyone changing any code. Old runs or entire sets of data can also be deleted on request, which shrinks the graph and removes them from future comparisons. The system is built around four core actions on the underlying memory graph: recording new information, recalling similar past cases, learning from human corrections, and forgetting data that should be removed. The idea the project is built to prove is that this kind of memory is what turns a one time judgment into something that improves over repeated use. Behind the scenes it uses a Python and FastAPI backend with a Next.js and React frontend dashboard, an AI model called DeepSeek to act as the judge, and a Jira integration that can read incidents and post back root cause comments and verdict tickets. It also connects to a tracing tool called Langfuse to record every judge call as a named trace. To run it locally you set up environment variables for the memory service, the AI judge, and Jira, then start the backend with a Python package manager and the frontend with npm. Sample scripts are provided that walk through recording a run, seeding history, and running a memory-aware evaluation.