DNA Memory

What DNA Memory does

DNA Memory treats a Markdown or Obsidian directory as the source of truth for long-term memory and SQLite as a deletable, rebuildable index. Clients connect over stdio MCP and use memory_recall, memory_remember and feedback tools, so they stop keeping separate and conflicting memories.

It does not store whole transcripts. Automatic import keeps only session IDs, paths, hashes, offsets and counts, and extraction produces at most three candidate memories per session, each up to 800 characters. Conclusions are sorted into eight types, including preference, fact, decision, workflow and error_lesson, and credentials, private keys and likely tokens are rejected. Recalled memories can be marked useful or misleading, and outdated ones are replaced through an explicit supersedes link. The profile config includes skill roots for Codex, Claude and Hermes, with Hermes at ~/.hermes/skills, and the README shows commands to connect all three clients to one MCP service. The README is in Chinese, with an English version linked.

Key features

• Markdown or Obsidian directory as the memory source of truth, SQLite as index
• Shared stdio MCP service for Codex, Claude Code, Claude Desktop and Hermes
• Eight memory types with duplicate detection and supersedes links
• Rejects credentials, private keys and likely tokens
• Stores session pointers instead of full transcripts

When to use it

• Letting Hermes and Codex recall the same project decisions
• Keeping long-term notes in an Obsidian vault that agents can search
• Marking recalled items as useful or misleading to improve later recall

Who it is for: People who switch between Hermes, Codex and Claude tools and want one memory instead of several.

How it fits with Hermes Agent

Hermes is one of the clients it connects, with a dedicated Hermes skill root in the profile; the tool also serves Codex and Claude.

Requirements: Python 3 with venv support; an MCP-capable client

FAQ