WYR-233/dsh-memory-lite

Read-only enhancer for a file-based long-term memory library (markdown cards kept in git) in DeepSeek Harness (DSH).

MIT · zero runtime dependencies · host + settings page · 中文说明见文末

Governance panels (0.6.x)

The settings panel now mirrors what the library scripts can do, and hides any block whose script is missing (it never assumes your library ships these tools):

Panel	Needs	What it shows
Vector fallback	memory.toml [vector]	fallback switch, vector DB / local model readiness
Search log	tools/search-log.mjs	query count, zero-hit rate, channel mix, top zero-hit phrasings (click to copy)
Regression gate	tools/memory-preflight.py	one button → rebuild index, run golden set, compare against baseline, per-row PASS/FAIL
Persona drift	tools/persona-drift.py	samples the replies it actually wrote in past sessions and scores caricature — it does not know it is being measured, so it cannot game the test

Card-size thresholds are read from memory.toml [card] (max_lines / max_bytes / soft_lines), so the panel always agrees with the library's own rules instead of a number baked into the plugin.

What it does

Feature	Detail
memory_find tool	Searches the library's cards: CJK by adjacent bigrams, Latin by whole words, weighted title ×6 / keywords ×4 / body ×1. Returns card paths + matched terms + a snippet, then the agent opens the card with read.
Memory digest injection	Appends one small systemPrompt section (default ≤900 chars, configurable): where the library is, which topics exist, and "search it before guessing".
Settings page	A "Memory library" section under DSH Settings: a governance dashboard (opening tax, scale & trend, layer health, consolidation, guard & backup, search tools), slider+number thresholds, seven one-click actions (checkup / backup / governance scan / new chat + prompt / prompt templates / in-panel search / open the memory site), and toggles (digest on-off, tool on-off, digest budget, search limit).

| Built-in search engine | The reference ranking (CJK bigrams, title ×6 / keywords ×4 / body ×1, synonym groups, idf weighting) is compiled into the plugin, so search works with zero setup; searchMode: 'script' still honours an external search script. | | One-click library | Pick a folder with the host's native directory picker, or create a library in one click: scaffolded INDEX.md, a sample topic folder, a sample card and a usage note. The chosen path lives in this machine's runtime config, so Windows and Linux each keep their own and nothing is typed by hand. | | /memory command + bundled skill | A slash command (status / search / init / root) plus a memory-library skill that teaches any agent the workflow — search first, write cards on real progress, keep the library small — so no preset is required. |

What it deliberately does NOT do

• Never writes to the library: no card edits, no commits, no pushes. For automation (auto-commit, periodic health checks) pair it with your own scheduled script.
• Never rewrites conversation history — it only appends one prompt section.
• No database and no cache; the only state is one small JSON config file.

Requirements

• A memory library directory with an index file (default INDEX.md) and a search script (default tools/memory-find.mjs). Any script speaking this contract works: node <script> --stdin --top <n> (query arrives on stdin) or --stats (library overview).
• DSH with @deepseek-ai/cordis resolvable in the profile (a normal DSH install).

Install

dsh plugin --profile web add github:WYR-233/dsh-memory-markdown

Then restart that profile.

60-second start (even if you have no memory library yet)

1. Install the plugin, then open Settings → Memory library.
2. Click Create a new library — or Choose an existing folder… if you already keep markdown notes somewhere. The path is written to this machine's own runtime config (<DSH_HOME>/memory-lite/config.json), so two operating systems never fight over one value, and the picker always yields a native absolute path.
3. That's it: the scaffold (INDEX.md, a sample topic, a sample card, a short usage note) is created and the built-in search engine is ready — no external script needed.

Prefer to let an agent do the shaping? Let an agent initialise it copies an initialisation prompt and opens a new chat, so the agent can ask what you want to remember and build the topic folders.

Where you will meet it

Where	What you get
Settings → Memory library	the dashboard: opening tax, scale & trend, layer health, consolidation, guard/backup, tools
Composer: /memory	status · search <words> · init [folder] · root <folder>
memory_find tool	the agent searches the library before guessing
Plugins page	a card for this bundle showing whether a library is configured
Below the composer	a one-line notice while no library is configured yet (dismissible)

Configure

Point it at your library — plugin config in the profile's cordis.patch.yml:

- id: memory-lite
  config:
    memoryRoot: /absolute/path/to/your-memory-library   # or set env DSH_MEMORY_ROOT
    stateDir: memory-lite                                # optional, relative to $DSH_HOME
    searchScript: tools/memory-find.mjs                  # optional
    indexFile: INDEX.md                                  # optional

or simply export:

DSH_MEMORY_ROOT=/absolute/path/to/your-memory-library

Runtime toggles (digest on/off, tool on/off, budget, search limit) are changed in the settings page and stored in <stateDir>/config.json — written atomically, and a parse failure falls back to defaults instead of silently wiping values.

Governance dashboard (0.3.0)

Beyond the digest and the search tool, the settings page is a read-mostly governance dashboard: six data blocks (opening tax of the always-read files vs a configurable budget, library scale and a 7-day trend snapshot, layer health with oversized-card and duplicate-title lists, consolidation status, git/guard/backup status, search-tool stats) plus six slider+number controls and a handful of one-click actions. Extra host routes:

Route	Purpose
GET /memory-lite/dashboard	all six blocks in one response (used by the panel)
GET /memory-lite/search?q=&top=	run the search script from the panel
POST /memory-lite/run	{"action":"checkup"｜"backup"｜"scan"} — runs the configured guard/backup script, or a pure in-host governance scan

Extra config keys (all optional, no path is hard-coded in the code):

- id: memory-lite
  config:
    taxFiles: [RULES.md, INDEX.md]          # files counted as the "opening tax"
    guardScript: ~/bin/my-guard.sh          # run by the checkup button
    backupScript: ~/bin/my-backup.sh        # run by the backup button
    backupDir: ~/backups                    # newest artifact is shown in the panel
    guardLog: ~/logs/my-guard.log           # last guard run
    coldDir: archive                        # cold area, excluded from search
    synonymFile: ''                         # optional synonym table (groups count shown)
    consolidationDir: ''                    # optional consolidation-card folder
    siteUrl: ''                             # optional library site; empty = no "open site" button

Every path above is optional: leave one empty and that block simply shows “—” in the panel. The plugin ships with no author paths, no domains and no uploads — it never talks to any server of its own, and the checkup/backup buttons only run scripts you configured yourself.

Thresholds (taxBudget, upkeepDays, upkeepCards, upkeepHardDays) live in the runtime config file and are editable from the panel — nothing is hard-coded in the reminder path either.

Optional: health panel

The Health card reads a log file when MEMORY_LITE_HEALTH_LOG points at one, expecting lines like:

[2026-01-01 10:00:00] issues=0 notes=2
  issue: <text>
  note: <text>

Any scheduled checker writing that shape shows up in the panel automatically.

How it fits together

This plugin is the thin "read fast" layer. The reference setup it was extracted from keeps memory as plain markdown cards in a git repo, plus an index of topics, a search script, optionally a periodic commit/health guard, and a rendered website for browsing the same cards. Because the store is plain files, everything stays human-readable, diffable and editable — that is the whole point.

中文说明

只读的记忆库增强插件：给 DSH 加一个 memory_find 检索工具、开场注入一小段记忆库摘要（默认 ≤900 字符， 可调），并在设置里加一栏「记忆库」侧边栏（概览 / 健康 / 开关）。

• 只读：不写卡片、不 commit、不 push（自动化请搭配你自己的定时脚本）
• 不改会话历史：只追加一段 systemPrompt
• 零依赖、无数据库：唯一状态是一个小配置文件

装好后把 memoryRoot 指到你的记忆库目录（或设环境变量 DSH_MEMORY_ROOT）即可；开关在设置页改， 即时生效，配置存在 <stateDir>/config.json（原子写，坏了回默认，不静默清空）。

License

MIT

---

开场协议文件（openingFile，0.5.0 起）

记忆库根下放一个 md（例如 OPENING.md），把「开场该读什么、报什么、有哪些硬线」写在里面， 插件会把它拼在摘要之前注入 system prompt。

- id: memory-lite
  name: 'dsh-memory-markdown'
  config:
    enabled: true
    openingFile: OPENING.md     # 相对 memoryRoot；留空 = 不注入
    openingBudget: 3000          # 开场协议字符上限（与摘要的 injectBudget 分开算）

为什么要有它：把开场协议从 agent 预设里搬进记忆库。

• 换预设不会崩（协议跟着库走，不跟着 persona 走）
• 改协议不用重启、不用动 YAML（每轮 assembly 重读文件）
• 协议本身是 markdown，可 git、可 diff、可回滚
• 文件缺失或为空时静默降级，不报错、不占位

取舍：仍然只注入一个 systemPrompt section（插件契约），opening 是拼进摘要的那一段； 它的长度受 openingBudget 管，不受 injectBudget（默认 900）限制。