@guowenzhang/dsh-memory
Claude 兼容的 DeepSeek Harness 记忆功能：用户选择的记忆目录，包含 MEMORY.md 索引和主题文件、记忆工具以及设置页面
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @guowenzhang/dsh-memory@0.1.2Memory 以 @guowenzhang/dsh-memory 发布，当前版本为 0.1.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 中文
DeepSeek Harness (dsh) is the open-source agent harness from DeepSeek AI, where nearly every capability is a plugin on Cordis. It is in developer preview and iterating fast, so expect compatibility-breaking changes (docs, 0.1.7-alpha.*); this plugin is a standalone third-party package that resolves @deepseek-ai/* from the running host.
A session forgot everything between runs; this plugin gives it a Claude-shaped memory directory (MEMORY.md plus one file per memory), a memory tool, and a settings section.
Screenshots pending — this plugin has no captured UI yet.
npx @deepseek-ai/dsh plugin --profile web add @guowenzhang/dsh-memory
From the npm registry: https://www.npmjs.com/package/@guowenzhang/dsh-memory — restart the host afterwards; local checkouts, git sources and troubleshooting are in AGENTS.md.
Settings → Memory starts with Enable memory, on by default. It is the master switch: off, no index is folded into new sessions and the memory tool is not registered, and nothing on disk is deleted.
<memory>/
├── MEMORY.md # Index: one line per memory, no frontmatter, loaded at session start
├── debugging.md # One memory per file, with frontmatter
└── ...
Index lines:
# Memory Index
## Project
- [build-and-test.md](build-and-test.md) — build ~45s, Vitest, dev server on 3001
- [architecture.md](architecture.md) — API client singleton, refresh-token auth
Topic files:
---
name: debugging-patterns
description: Auth token rotation and database connection troubleshooting
metadata:
type: reference
---
## Auth tokens
...
The Memory directory template field decides where the store lives; ~ is your user home and {project} is the current project's directory name.
| Form | Resolves to |
|---|---|
~/.dsh/memory/{project} (default) | One subdirectory per project, so projects never share memories |
~/.dsh/memory | One store shared by every project (drop {project} to share) |
~/.claude/projects/{project}/memory | The same store the Use Claude's directory switch selects |
{project} is the git repository root of the session's working directory (a .git file — a linked worktree — is resolved back to the main repository), with every character outside [A-Za-z0-9] replaced by -: C:\02-codespace\deepseek-harness becomes C--02-codespace-deepseek-harness, exactly the directory name Claude writes under ~/.claude/projects/.
Choose directory… opens the OS picker when the host composes one and the in-app directory browser otherwise; typing a path into the field works everywhere. Beside it, Directory state reports the directory the current project resolves to, the index's lines and bytes, the file count and total size, and the file cap.
Use Claude's directory is off by default. While it is off, memory lives in your own directory, which you choose and can change at any time; while it is on, there is no directory to choose — the store is Claude Code's ~/.claude/projects/<project>/memory, shared with Claude itself. Turning it back off returns you to the directory you had configured.
Memory is stored per project, so the page shows one at a time: the Directory state block and the MEMORY.md index editor describe whichever project the Project dropdown selected. The list is
<claude home>/projects/.Rows are deduplicated by the directory they resolve to, so one repository appears once, and a single-project deployment shows no dropdown at all. Selecting another project only changes what the page reads and edits: the template is untouched, and every session still resolves its own directory.
Saving takes two steps: write the memory file, then add one pointer line to MEMORY.md — the index holds pointers, not content. The injected instructions state that protocol and the model follows it; you can also edit the index directly in the settings page's MEMORY.md index editor.
memory toolThe model reaches the whole store through one memory tool whose paths are addressed under /memories, and every reply is the tool's own wording — including the failures, which are returned rather than thrown (Error: File … already exists, Please ensure it is unique, …). The view line ceiling, view_range, the two-level directory listing, the hidden-item and node_modules exclusions, /memories addressing, create failing on an existing file, the unique old_str requirement, the 0-based insert_line, delete refusing the root, and rename refusing an existing destination all match Claude's memory tool.
| Command | Behavior |
|---|---|
view | Lists a directory two levels deep, hidden entries and node_modules excluded, or shows a file with line numbers; view_range: [start, end] windows a file |
create | Writes a new file, creating parent directories; fails when the path already exists |
str_replace | Replaces an old_str that must occur exactly once; reports the line numbers when it occurs more than once |
insert | Inserts text at the 0-based insert_line |
delete | Removes a file, or a directory and its contents; refuses the /memories root |
rename | Moves an entry; refuses an existing destination |
The first request of every session folds one instructions message: the memory protocol plus the current MEMORY.md, the first 200 lines or 25 KB — the same bound Claude loads by. The message is recorded on the Session log, so a resumed session is not given it twice; the bound is a load limit, not a trim, and the rest of the index stays on disk, out of the request.
modified frontmatter stamp. Claude Code adds an ISO-8601 timestamp when it writes through its normal file tools; this plugin never rewrites what the model wrote.# append shortcut: current official Claude documentation has none either (interactive mode lists only /, !, @, :, ?). To record something by hand, edit the index in the settings page or have the model use the memory tool.memory tool, and a deletion only when the model runs delete or you do it from the settings page.../, absolute paths, and drive-qualified paths never escape /memories, and a symlink inside the directory cannot redirect a write outside it.CLAUDE.md instruction files are not this plugin's surface — that is @guowenzhang/dsh-claude-compat. The two are independent and install separately.The plugin itself is Apache-2.0 — see LICENSE and NOTICE. Portions of the plugin contract, the settings section, and the Remote surface are derived from DeepSeek Harness and remain under the MIT License reproduced in NOTICE.
CLAUDE.md instruction files, skills, and scope rules.