When to Forget: A Memory Governance Primitive

The project addresses a critical bottleneck in agentic systems: the 'clutter' of long-term memory where irrelevant or misleading historical data degrades performance. By proposing a 'Memory Worth' metric based on outcome feedback rather than static importance, it introduces a principled way to perform credit assignment on retrieved context. However, the defensibility is low (3/10) because this is a 'primitive'—an algorithmic idea rather than a moat-driven product. With 0 stars and 1 fork, it currently lacks the community momentum or network effects found in projects like Letta (formerly MemGPT) or Zep. Frontier labs like OpenAI and Anthropic are already experimenting with personalized memory and context management; they are highly likely to implement similar outcome-based pruning natively at the platform level. The primary value here is as a reference implementation for agent framework developers (LangChain, CrewAI) to incorporate into their middleware, but as a standalone entity, it faces significant 'feature-not-product' risk.