dsh-md-memory
Manifest validSmall file memory for DeepSeek Harness. The model writes markdown, the host injects it.
dsh-md-memory
English | 简体中文
A small file memory plugin for DeepSeek Harness. The model writes what should persist as markdown under ~/.dsh/memory. Next session, the host injects it.
This is not a second brain. The community already has fuller memory plugins (auto-extract, vector search, settings dashboards). This one stays small: the disk holds files you can open in a text editor, and nothing is saved unless the model calls memory.
MIT
What it does
- Remembers preferences, corrections, and reusable lessons across sessions
- Three files:
user.md(cross-project) /agent/MEMORY.md(hot rules) /agent/topics/*.md(on-demand topics) - Project-only facts belong in the repo
AGENTS.md, not here - Subagents cannot write, so parallel tasks cannot scramble memory
- If you say things like “remember / from now on / don’t ever”, and nothing was stored this turn, the host nudges the model
What it does not do
- No automatic extraction from the chat
- No secret scanning (do not let the model write tokens here)
- No vector / full-text engine (
searchis a plain substring) - No browser settings page
Do not install this next to another plugin that also registers a tool named memory.
Install
dsh plugin --profile web add github:CuteSamurai24/dsh-md-memory
Then restart dsh web.
Uninstall:
dsh plugin --profile web remove dsh-md-memory
Files stay in ~/.dsh/memory. Uninstalling the plugin does not delete them.
Where the files live
~/.dsh/memory/
user.md You, across projects: identity, tone, standing preferences
agent/MEMORY.md Hot rules for almost every session (keep this short)
agent/topics/*.md Topics, read when the description matches
On Windows that is C:\Users\<you>\.dsh\memory\. Open the files in Notepad to check, edit, or delete.
How the model writes
One tool, named memory:
| Operation | Use |
|---|---|
read / search / list | Read |
append / edit / write | Write a layer |
create / delete | Manage topics |
Appending to user requires a one-sentence reason: why this still holds on a different project.
Caps: user 8KB, hot rules 12KB, one topic 20KB. Long material goes in a topic.
Turn it off
On by default. In ~/.dsh/settings.yaml:
md-memory:
enabled: false
Later turns in this process stop injecting and stop offering the tool. Restarting dsh web is the simplest way.
Other memory plugins
DSH has no official long-term memory product. The community already has KV stores, auto-extract, persona files, and heavier “evolve” suites. This plugin only covers one narrow need: the model writes markdown, a human can open it.
License
MIT © CuteSamurai24
Comments
Loading…
Similar plugins
by SYMlp
Markdown folder long-term memory plugin for DeepSeek Harness (dsh): one file per fact, human-editable, git-versionable
★ 0
MIT
TypeScript
Aug 20, 2026
dsh plugin --profile web add dsh-markdown-memoryLayered file memory for DeepSeek Harness with workspace-scoped USER/MEMORY notes, background consolidation, and hybrid session recall.
★ 0
↓ 55/wk
dsh plugin --profile web add dsh-file-memoryby YYTbit
Structured long-term memory system for DeepSeek Harness
★ 5
↓ 33/wk
MIT
TypeScript
Aug 14, 2026
dsh plugin --profile web add dsh-plugin-meta-memoryby wojiao42
文件型长期记忆:每个会话注入你自己的 Markdown 记忆文件,提供读写工具,并能查出 DSH_HOME 漂移;不读对话、不联网 · File-backed memory for DeepSeek Harness, with verifiable constraints
★ 0
MIT
JavaScript
Oct 7, 2026
dsh plugin --profile web add dsh-memory-fileby disc0nct
A persistent memory plugin for DeepSeek Harness that enables agents to store and recall information across sessions — DSH-native tools with atomic storage, validation, and concurrency-safe reads
★ 0
MIT
JavaScript
Aug 22, 2026
dsh plugin --profile web add @deepseek-ai/dsh-tool-memoryby guyuefangyuanl
Cross-session persistent memory for the DeepSeek Harness: a model-facing memory tool, an always-on index section, and a bundled maintenance skill.
★ 3
↓ 370/wk
MIT
JavaScript
Aug 21, 2026
dsh plugin --profile web add dsh-memory