rawmem + memdsl: raw evidence and reviewed agent memory over MCP

[Liyuan1992]

Pre-submission Checklist

• This post follows the MCP community guidelines

What would you like to share?

I built two small open-source MCP servers for agent memory:

• rawmem keeps a local, append-only record of past work, so an agent can look up what happened before. It exposes bounded summaries by default.
• memdsl keeps long-term rules, preferences, and decisions only after review, so unapproved notes do not silently become memory.

Used together, rawmem keeps the evidence and memdsl keeps what you have approved.

They work over stdio MCP with Codex, Claude Code, and DeepSeek Harness. Both projects include setup guides, demos, and tests.

I'd love feedback on setup and real-world use cases.

Relevant Links

rawmem: https://github.com/Liyuan1992/rawmem
memdsl: https://github.com/Liyuan1992/memdsl