kedarvartak/memops

MemOps is a monitoring and debugging toolkit for AI agents that use persistent memory, the kind of memory that lets an AI remember facts across multiple conversations or sessions. The problem it addresses is that persistent-memory agents can degrade in subtle ways that ordinary infrastructure dashboards do not catch: the agent might return stale or contradictory facts, repeatedly store near-identical memories, or slow down as the memory layer grows. MemOps captures a stream of memory events from an agent, including writes, reads, and rewrites, stores them in a local SQLite database, and computes health metrics from those events. Metrics include churn rate, or how often memories are rewritten, duplicate rate, contradiction score, stale recall rate, empty retrieval rate, and average retrieval latency. These turn invisible memory failures into numbers you can inspect and threshold. The regression analysis feature lets you compare two runs, a known good baseline against a candidate run after a code or policy change, and get a delta report showing which metrics worsened. The tool can return a non-zero exit code, making it usable inside automated test pipelines and continuous integration workflows. Sessions can be replayed step by step to pinpoint exactly which turn introduced a contradiction or loaded a stale memory. A local dashboard surfaces session timelines and root cause hints. The project is currently command-line first and built primarily around the Letta agent framework, though its internal event architecture is designed to support other frameworks through adapters. It is written in Python 3.11 or later.