A memory layer for LLMs that begs you not to ship it

What it does Kernel Memory² is Microsoft’s second research prototype for AI memory, described only as a “Memory solution for users, teams, and applications.” The README is deliberately thin on mechanics, stating that both the old and new codebases exist purely to explore new ideas and gather community feedback.

The interesting bit The rewrite is informed by lessons from the first iteration, and the team is using concepts from Amplifier—their parallel metacognitive AI engineering platform—to construct the next generation. Future work is slated to focus on content quality, privacy, and collaboration, though the README admits details will come “when ready.”

Key highlights

• Full rewrite of the original prototype; the previous codebase remains in the repo strictly for reference
• Being built with AI and concepts from Amplifier, a platform for metacognitive AI engineering
• Future roadmap lists content quality, privacy, and collaboration as focus areas
• MIT licensed yet explicitly framed as a learning resource, not production-ready software

Caveats

• The README is vague on architecture, APIs, and capabilities; most technical specifics are undescribed
• Contributions are not accepted, no support is offered, and the team warns to “expect things to break”
• No stability or compatibility guarantees; the maintainers suggest pinning specific commits if you need consistency

Verdict Worth watching if you track Microsoft’s AI incubations or research into memory and RAG patterns. Avoid if you need a stable, documented backend for shipping applications today.