dsh-soul-md

The plugin follows the DSH language setting (Chinese and English in DSH 0.2.0-rc.2), including configuration, status messages and plugin-list metadata. Language-pack locales use the host fallback chain. Switching languages preserves unsaved settings; there is no separate plugin language selector.

v0.9.1: resident memory aliases and bounded index titles

• All three memory_* tools treat topic: "core" / "core.md" (case-insensitive, surrounding whitespace ignored) as an omitted topic, reading/writing resident core memory. Other topics keep their existing behavior.
• Index titles are capped at 80 characters and summaries at 240; topic keys remain distinct and full files are unchanged.
• Existing topics/core.md files are preserved without automatic merging. Back them up, inspect them with a file reader, merge resident entries into core.md, and rename remaining topic content to another key. New memory_read({ topic: "core" }) calls read resident core.

Keyword search and optional recall

With layered memory enabled, memory_search({ query: "Electron packaging", limit: 5 }) searches topic keys, titles, summaries and full bodies, returning keys and summaries only. Persona topics override global topics with the same key; missing keys fall back to global topics, just like memory_read(topic). English matching is case insensitive; Chinese phrases also use overlapping two-character terms. This is lexical search, so synonyms without shared text require another query. Search does not require automatic recall to be enabled.

The configuration page offers these independent, default-off options:

• memory.recall: "keyword": from the second human message handled by the live plugin, add a soul:recall section with possibly relevant keys and summaries. The first message and internal context/compaction summaries do not trigger it.
• memory.indexMode: "recall": while keyword recall is enabled, replace the full injected topic index with a short search hint. The existing index generation, title/summary bounds, truncation rules and memory_read() index stay unchanged.
• memory.compactionRecall: true: after successful compaction or pruning, the next human message restores descriptors for topics read/appended/rewritten in this session. They remain stable for that turn, and clear on the next human turn. Failed compaction does not trigger recovery. Recovery tracks at most the latest 128 topic keys in memory; unloading/restarting the plugin resets state.

Recall and recovery never inject topic bodies. They share recallMaxChars (default 1600, maximum 8000) and the remaining injectMaxChars budget after the existing memory section; if core/index uses the budget, no extra rows are added. Each group includes at most recallMaxTopics (default 5, maximum 20). Dynamic descriptors always preserve literal braces, even with allowTemplates. memory.inject: false and skipSubagents also suppress soul:recall.

Suggested opt-in configuration in the soul-md row:

memory:
  layered: true
  recall: keyword
  indexMode: recall
  compactionRecall: true
  recallMaxTopics: 5
  recallMaxChars: 1600
  injectMaxChars: 24000

Configuration page (DSH 0.2.0-rc.2 and later)

Open Plugins → Installed → dsh-soul-md from the homepage sidebar to configure and save this plugin. The page uses the official plugins.bundle.config interface, without a duplicate entry in global Settings. Web and Desktop share the page. This version requires DSH 0.2.0-rc.2 or a later 0.2.x host; existing configuration is retained.

GitHub: Scorp1o117/dsh-soul-md · npm: dsh-soul-md

Part of the DeepSeek Harness Enhancement Suite — Vision · Soul/Persona · Long-term Memory · Plugin Marketplace.

Persona + long-term memory for DeepSeek Harness — zero file management:

In Settings → 人设卡, type a card name and its content, hit save. The plugin manages everything else.

What you get

• Persona cards — the card content is rendered into the system prompt as the soul:persona section. Multiple cards are supported; pick a default, and switch per chat from the conversation header (a "人设" select).
• Long-term memory — the agent gets six tools:
  • memory_append / memory_read / memory_rewrite — a persistent memory file (Agent.md / memory.md style). The active persona card has its own memory; otherwise the global memory is used. In layered mode, their optional topic argument reads or writes one on-demand topic.
  • memory_search — search layered topic content and return keys and summaries.
  • soul_read / soul_update — the AI reads and evolves its own persona card: when it notices a stable trait, preference, or value of its own, it folds it into the card. It "grows" across sessions instead of resetting every time.
  • The memory is also injected as a soul:memory prompt section (capped) so the agent always sees its memories.
• Resolution per prompt assembly: session choice (chat switcher) > workspace mapping > default card > none. Switching applies from the next turn — no restart.
• Workspace personas (v0.5.2): Settings → 人设卡 lists every workspace with a card dropdown — sessions of that workspace use the assigned card by default (session-level switching still wins). Workspaces come from dsh's durable workspace registry, so no paths to type.

Desktop install

Use the Desktop-installed dsh command (Application → Manage dsh Command), or the app’s Plugins page. Then install into the Desktop profile:

dsh plugin --profile desktop add dsh-soul-md

Restart the Desktop app to load the client bundle. Desktop keeps its profile under $DSH_HOME/profiles/desktop.

If Hub reports Cannot find module '...app.asar/dsh/node_modules/@deepseek-ai/dsh-desktop-host/lib/cli.js', the Desktop CLI bootstrap failed before loading this plugin. Reinstall a complete official Desktop distribution, or try the official Plugins page (outside Hub). Switching this command to the Web profile does not repair the Desktop package. See issue #14.

Install

The plugin is a plain Cordis row. Mount it in a profile patch ($DSH_HOME/profiles/<name>/cordis.patch.yml):

- insert:
    - id: soul-md
      name: 'dsh-soul-md'          # after: pnpm add dsh-soul-md in the profile

Then restart dsh web and open Settings → 人设卡: type a name + content, save.

Where things live (you don't need to care, but for reference)

• Persona cards: stored in the soul-md entry of the active Profile patch, as cards: { name -> markdown } + active + per-session sessions.
• Memory files: plugin-managed under $DSH_HOME/soul-md/memory/ (global.md + one file per card), created on demand.
• Optional layered memory (off by default): memory/<card>/core.md is injected within the configured cap, while memory/<card>/topics/*.md contributes its title and first body line to an index. memory_read({ topic: "..." }) retrieves a topic's full text. Existing <card>.md files remain readable and seed core.md on the first layered append, so enabling the option does not hide old memory.
• Upgrading from ≤ v0.4 (file-based)? The plugin auto-imports the old path card (as "默认") and the old memory file on first run.

Config

skipSubagents is off by default and keeps the v0.7.0 behavior. When enabled, sessions that DSH created as delegated children (origin: "subagent") no longer receive the soul:persona / soul:memory sections — a child usually does one small job and does not need to carry the parent's full persona and long-term memory every turn.

It only affects prompt injection: cardNameOf, memoryTarget, and the scope of the three memory_* tools are unchanged, so a child reads and writes exactly the same card and memory files it would otherwise (nothing silently falls back to global.md).

Compatibility: tested with @deepseek-ai/dsh@0.1.7-rc.1 and 0.1.7-rc.2 (npm next); npm latest is 0.1.5-rc.3. This version uses Profile patch settings and browser configForms. Older hosts require an older plugin release. Alpha builds remain unknown.

v0.8.1: reliable persona switching and layered memory truncation

• Reports refused session persona writes, rolls the selector back to the Host snapshot, and blocks overlapping selector changes while a write is pending.
• Keeps the topic index when layered memory exceeds its injection cap, retains both ends of core, and names omitted content in the prompt.
• Passed install, Web boot, homepage and client-bundle HTTP checks, and uninstall in a disposable DSH 0.1.5-rc.3 Profile.

v0.8.0: subagent prompt control and DSH next compatibility

• Adds opt-in skipSubagents: delegated child sessions (origin: "subagent") no longer receive the soul:persona / soul:memory sections.
• Render path only; cardNameOf, memoryTarget, and the memory_* tool scopes are unchanged.
• The settings page exposes the switch in the long-term memory group, saved together with memory.inject / memory.layered.
• Records DSH 0.1.5-rc.3 (next) compatibility; unverified alpha releases remain unknown.

v0.7.0: Layered memory

• Adds opt-in memory.layered; existing users keep the single-file behavior.
• Injects core.md as resident memory and only a title/summary index for topics/*.md.
• Adds an optional topic argument to all three memory_* tools for on-demand topic reads and writes.
• Keeps legacy <card>.md visible as the core fallback and carries it into the first layered append.
• Passes disposable-profile install, Web boot, client-bundle, and uninstall smoke checks on DSH 0.1.5-rc.2.

Notes

• Since v0.8.5, braces stay literal by default in persona cards and memory. If an existing card relies on host variables such as {{cwd}}, enable Allow prompt variables in persona and memory (allowTemplates: true) in settings. Unknown variables still fail rendering in that mode.
• Persona/memory sections resolve per assembly, so steady cards stay byte-identical (KV-cache friendly) and edits hot-apply.
• DSH exposes the registered soul-md settings namespace directly; the plugin does not modify files in the host installation.
• Suggest putting work-quality rules in the card (e.g. "task quality first") so roleplay never degrades real work.
• Version 0.5.8 and newer require DSH 0.1.0-rc.7 or newer and are tested against 0.1.0-rc.7, 0.1.0-rc.8, and 0.1.1-rc.1.
• DSH 0.1.0-rc.6 users must pin dsh-soul-md@0.5.6, the last release carrying the legacy settings-allowlist compatibility patch.

License

MIT