Memory with a visible paper trail for DeepSeek Harness
dsh-mnemon gives DSH agents layered, inspectable memory — plus skills that draft their own improvements and backtest them before you adopt anything.

What it does
It’s a DSH plugin that splits agent memory into three tiers: runtime memory (preferences and working facts, injected every turn as compact USER.md/MEMORY.md), Project Documents (searched on demand), and Memory Spaces (durable facts and entity relationships recalled from pluggable providers). A “turn memory bar” under each reply shows exactly which documents and memories that turn read and wrote, one click from opening them. Saving a fact is a brain-mark click that hands it to a task agent for deduplication and distillation.
The interesting bit
The Skills component is the unusual part: it reviews your real finished conversations in the background, drafts improved SKILL.md files, and backtests the draft against the current version before you adopt it — with a diff and a score. Manual adoption is the default; automatic adoption is opt-in and only after non-regressing backtests.
Key highlights
- Two strategies: Layered (runtime resident, rest on demand) and General (everything within one budget, model decides)
- Memory System UI: entity graph linking memories, an Ask Agent that answers with citations, and full editing of stored memory
- Local-first via Mnemon Native, or eight third-party providers; storage can be global, per-workspace, or a directory you pick, with ZIP backup
- Sources and strategies are ordinary DSH plugins on public SDKs — extensions get the same settings pages as shipped components
- Skill market installs from skills.sh, GitHub, and any Agent Skills discovery index, with previews and security reviews where available
Caveats
- Skills is explicitly experimental; self-evolution is off by default
- Memory Spaces need a separate CLI install (Mnemon Native) or a third-party provider — not functional out of the box
- DSH itself is on RC releases, with version-dependent Node.js requirements (up to >=24.9.0)
Verdict Worth a look if you run DeepSeek Harness and are tired of agents with amnesia — the per-turn memory audit trail alone is a genuinely good idea. If you’re not in the DSH ecosystem, it’s a design worth studying, not a tool to install.
Frequently asked
- What is omdsh-dev/dsh-mnemon?
- dsh-mnemon gives DSH agents layered, inspectable memory — plus skills that draft their own improvements and backtest them before you adopt anything.
- Is dsh-mnemon open source?
- Yes — omdsh-dev/dsh-mnemon is open source, released under the MIT license.
- What language is dsh-mnemon written in?
- omdsh-dev/dsh-mnemon is primarily written in TypeScript.
- How popular is dsh-mnemon?
- omdsh-dev/dsh-mnemon has 1.1k stars on GitHub.
- Where can I find dsh-mnemon?
- omdsh-dev/dsh-mnemon is on GitHub at https://github.com/omdsh-dev/dsh-mnemon.