dsh-simple-memory 🧠

English | 简体中文

A simple memory keeper plugin for DeepSeek Harness (dsh) web: retrieval-style memory with zero-code storage. No database, no vector store — just well-organized markdown files plus a thin plugin that handles the entry points. Normal work is never interrupted; memory accumulates as a sidecar filesystem beside your projects, and rules nudge you to record a note from time to time.

Status: core pipeline verified (0.2.0), real-world effect still under testing — see docs/verification-log for the test run and docs/design for how it works.

Unofficial project: independently developed and maintained by a community member, not an official DeepSeek product.

Screenshot

Memory button (the bulb icon left of the input box): one click starts the memory flow — review → propose each item with a project/global scope and reason → confirm → write → show output → commit:

Memory management page (Settings → Memory): status overview, memory-root config, flat per-project lists plus global, click to read:

Features

Install

dsh plugin --profile web add "github:a903067276-rgb/dsh-simple-memory#main"

Restart dsh web, then Settings → Memory → Initialize memory repo. Manual install fallback: see docs/install.md.

Usage

• Memory button (the bulb icon left of the input box) — one click inserts a fixed, self-contained memory-flow instruction; the model then runs the standardized flow (review → propose each item with a project/global scope and reason → confirm → write → show output → commit).
• Settings → Memory — status overview, memory-root config, one-click repo init, and an inline browser: flat per-project lists plus global, click to read.

Platform support

Requirements

• DSH web
• git CLI (optional: without git, the memory repo is just a plain directory)

How it works

A retrieval-style memory: files are the storage, the plugin only handles the entry points. Full design: docs/记忆系统设计说明.

• Storage (zero code) — everything lives in one memory root ~/Documents/DSH/memory/ (configurable from the settings page):

  memory/
  ├── common/          global experience (active zone, cross-project reuse)
  ├── projects/<name>/ per-project memory (auto-created on first write)
  ├── references/      cold zone: reference material (search-only)
  ├── archive/         cold zone: forgotten memory moved out of the index
  ├── staging.md       promotion pool (candidates for global reuse)
  └── .git/            one git repo for the whole root — rollback safety

  Per-project memory does NOT live inside the project directory: a .gitignore'd memory/ would hide it from grep. Keeping it in the shared root keeps publication isolation automatic and search universal.

• Index injection (remembers what exists) — on the first step of every session (agent/pre-step), the plugin injects a filename list (never the content), grouped by category, covering the current project + global + loose root files. The model then reads a full note (≤2KB each) on demand. Cold zones are not injected — searched only when a topic hits.

• Write tool (records it) — memory-write enforces the format: filename 分类-主题.md, first line ## date 分类-主题, ≤2KB, category prefixes open (built-in: 踩坑/流程/决策/偏好/背景). scope picks project or global. Writing outside the workspace asks for approval (built-in confirmation).

• Memory button (triggers the flow) — the bulb icon in the input bar inserts a fixed, self-contained instruction: ① review the conversation ② propose each item with a scope (project/global) + reason ③ wait for user confirmation ④ write via memory-write ⑤ show the output ⑥ git commit (mem: 记 xxx).

• Settings page (manage + browse) — status line (active count, staging count), memory-root config, one-click repo skeleton init, and an inline browser listing every project flat plus global, click to read.

• Search (finds it) — no index files: the agent's grep scans the whole memory root (all projects + global + cold zones) in one pass, active zones first.

• Promotion (cross-project reuse) — project memory that looks reusable goes into staging.md (low friction, no instant decision); when the pool is non-empty the user is reminded, and after confirmation it is distilled into common/ and removed from the pool.

• Forgetting (fresh context) — outdated notes move to archive/ (soft delete: still on disk, just out of the index). Context stays lean, disk stays complete.

• Rollback — every memory operation is committed immediately (mem: <action> <subject>); a wrong write is one git checkout away.

Notes

• Memory git repos are local-only, never pushed to a remote.
• Writing memory outside the workspace asks for approval (built-in confirmation, per the memory spec).
• The memory-flow instruction is self-contained and does not depend on the global AGENTS.md; the judgment rules there are only a reference.

License

MIT