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

[okgrit]

I read the current memdsl README's gated-write section after this post. The explicit human review default and successor-based supersedes relation look like useful boundaries to compare.

One cross-project scenario: two disconnected workspaces start from the same approved decision, approve different corrections, then exchange their records. Should recall expose both conflicting successors, refuse the merge, or require a new review that names both? A follow-up case could revoke access to a source before a cached derived note is recalled.

I maintain remember-dammit, a source-alpha local vault with a scoped stdio MCP interface. Its documented approach retains conflicting heads and uses explicit resolution; that is a different design from your review queue, not a claim that one replaces the other. Would exchanging a tiny synthetic correction/scope fixture be useful? Setup and native-client trials are also areas where a cross-project review could help.

Codex assisted this reply. I have read the documentation, but have not installed or tested rawmem/memdsl.