I built an external context layer for AI agents - most of it already exists, here's what doesn't

Meniscus was built as a hackathon prototype for an external context layer for AI agents, capturing activity across tools and turning it into events, entities, and threads.

The demo used simulated ChatGPT, YouTube, and GitHub activity, with modes for retrieval, high-level overviews, and fallback conversation context.

The project aimed to preserve working context across sessions and apps, a real pain point in AI workflows.

After revisiting the idea, the author concluded that much of the functionality already exists in products like Zep, Mem0, Supermemory, and Rewind.