mindspace-dsh-session-memory
适用于 DeepSeek Harness 的任务条件式 Chat 和 Work 会话记忆
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Spirtxiaoqi7/mindspace-dsh-session-memory#117df92bba55c9e17da4e8b858cad4dce16214a4Mindspace Dsh Session Memory 以 mindspace-dsh-session-memory 发布，当前版本为 0.7.0。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。
中文 · Community plugin · MIT
Separate work from everyday conversation. Maintain memory beyond context summaries.
Chat and Work are two memory contexts for the same user and assistant, not two tool-permission presets. Each stores people, relationships, preferences, assistant identity, current state and lasting experiences.
Use the composer Chat / Work chip to express the current context. The model may also switch when the main intent changes. Modes never disable tools or alter permissions.
In Settings → Personalization, inspect/edit memory, inherit it into a new session, change the compaction policy or compact manually.
Cross-context facts are staged in a neutral bridge: a short transition note (up to 300 characters) plus pending write instructions. Only after entering the target mode does the model merge or dismiss each pending item.
Compaction captures the original textual conversation since the last successful compaction. On success, a separate request receives this evidence and current memory, then proposes source-linked operations. Long conversations are processed in ordered batches rather than truncated.
Memory revisions are checked before applying changes. Concurrent edits cause a retry against the latest state; cross-mode changes remain staged. Failed calls preserve memory and task evidence, with up to three attempts per batch. Background maintenance does not block ordinary conversation.
This incurs additional model usage after compaction, not after every turn. Empty provider/model settings reuse the session route; a dedicated lower-cost model can be selected. No tools are supplied to the maintenance call. Model judgments can still be corrected in the Memory Center.
Override the installed bundle in your Web profile's cordis.patch.yml:
- id: mindspace-session-memory
config:
maxTextBytes: 4096
maxItemsPerSection: 100
maxProfileCharacters: 300
maintenanceEnabled: true
maintenanceProvider: ''
maintenanceModel: ''
maintenanceMaxTokens: 6000
Restart after configuration changes. Compaction enablement/thresholds are per-session; maintenance configuration is plugin-wide. Successful manual compaction also triggers maintenance.
Version 0.7 targets DSH 0.1.5-rc.2 and the corresponding 0.1.x APIs. Keep plugin 0.6.10 on the older 0.1.1 core until upgrading DSH.
Storage remains under DSH_HOME/mindspace-session-memory/v1; the directory name is not the document format version. V5 Chat/Work documents remain unchanged. Existing V1–V4 data is read through the migration path into Chat without inventing a Work persona. Legacy memory events can still be imported; new memory writes remain outside canonical session logs.
Maintenance tasks and original evidence live in DSH_HOME/mindspace-session-memory/maintenance. These are local user data, not repository artifacts. Back up DSH_HOME before upgrading. DSH handles its own session-format migration; use the matching pre-upgrade backup when rolling back.
git clone https://github.com/Spirtxiaoqi7/mindspace-dsh-session-memory.git
Set-Location .\mindspace-dsh-session-memory
corepack pnpm install
corepack pnpm run check
$memoryTgz = (Get-Item .\dist\mindspace-dsh-session-memory-0.7.0.tgz).FullName
Set-Location C:\path\to\deepseek-harness
corepack pnpm dsh plugin --profile web add $memoryTgz
corepack pnpm dsh web
No DSH core patch is required. Do not load a duplicate legacy implementation alongside this plugin. Manual edits, model writes and maintenance share the same storage and revision path.
See CHANGELOG. This is not an official DeepSeek project.