dsh-memory
Global memory plugin for the DeepSeek Harness (dsh). It gives every agent a persistent, cross-session long-term memory: durable facts, user preferences, and decisions that survive across sessions, projects, and workspaces — and it automatically recalls them into every task.
> 📖 Full usage tutorial (install, verify, daily use, config, FAQ): [docs/tutorial.md](./docs/tutorial.md) (中文)
> 中文说明见 README.zh.md。
What it does
| Feature | How it works |
|---|---|
| Global store | Memories are plain Markdown files under $DSH_HOME/memory/*.md (shared by every session, not tied to any workspace). |
| Automatic recall | The memory context is injected into every agent step, so every task starts with your memories already loaded. |
| Automatic input | The agent saves memories itself through tools — no manual file editing. |
| Auto-summarize | memory_summarize distills a long input (or a recap of the conversation) into a compact memory using the language model. |
| Human commands | /memory lists (or searches) memories from the chat UI. |
Model-facing tools
memory_add— save a durable fact / preference / decision. Reusing the same title (orid) updates the existing memory.memory_search— keyword search over titles, tags, and body.memory_list— list all memories (optionally filtered by tag).memory_delete— forget a memory by id.memory_summarize— distill text into a compact memory via the LLM, then save it.
The standing injected instruction also tells the agent to proactively keep memory current (save new durable facts, update stale ones, avoid duplicates).
Install
With the GitHub CLI and pnpm installed, add the plugin to a profile:
# git source (this repository)
dsh plugin --profile web add github:lifensame/dsh-plugin
# or, from a local checkout
dsh plugin --profile web add <path/to/dsh-plugin>
# or, once published, from npm
dsh plugin --profile web add dsh-memoryThen restart the profile. The first time you add it, pnpm installs the package and dsh reconciles it into the profile's bundle list automatically.
How memories are stored
Each memory is one Markdown file with YAML frontmatter:
---
id: favorite-color-abc123
title: Favorite color
tags: [prefs]
importance: high
created: 2026-01-01T00:00:00.000Z
updated: 2026-01-02T00:00:00.000Z
---
The user's favorite color is blue.They live under $DSH_HOME/memory (~/.dsh/memory by default), so you can also edit them by hand or put them under version control.
Configuration
All options are optional; the values below are the defaults.
# profile cordis.patch.yml — the dsh-memory row
- id: memory
name: dsh-memory
config:
memoryDir: memory # directory (relative to $DSH_HOME) holding the .md files
maxInjectedBytes: 16384 # UTF-8 budget for the auto-injected memory context
searchDefaultLimit: 5 # default result count for memory_search
summarizeProvider: '' # explicit LLM provider route for memory_summarize (empty = default model)
summarizeModel: '' # explicit LLM model id, paired with summarizeProvider
summarizeMaxInputBytes: 24576 # max input size memory_summarize accepts
summarizeMaxOutputTokens: 512 # max tokens for the summarization call
summarizeTimeoutMs: 60000 # timeout for the summarization callQ-version logo
A ready-to-use prompt for generating a chibi (Q版) DeepSeek-memory mascot is in [docs/image-prompt.md](./docs/image-prompt.md).
License
MIT