跳过主要内容
K

dsh-memory-switch

kingzhoudk/dsh-memory-switch

按会话切断跨会话记忆通道的开关:关闭后跨会话历史搜索与追溯工具及共享的研究缓存被拒绝,跨会话 @ 引用与团队转交消息被过滤,安装 OpenViking 时其回忆、捕获与提交也被拦截;子代理继承状态,刷新、恢复与分叉都会从会话日志还原,默认开启。

安装

dsh plugin --profile web add github:kingzhoudk/dsh-memory-switch

README

dsh-memory-switch — per-conversation memory switch

One switch per conversation. Turned off, that conversation stops exchanging information with other conversations through any of the channels listed below.

It is a policy gate, not a sandbox: it closes the seams a plugin can see (tools, injected messages, wrapped services). Reading ~/.dsh/sessions/ directly through bash / read needs a sandbox policy — see Known limits.

  • The switch lives in the composer tool row (conversation.input.left, beside the permission / plan chips) and is per conversation.
  • State is derived entirely from the session log: clicking the switch runs /memory on or /memory off, the host folds that command into the memorySwitch session projection, and the projection frame pushes the new state back. Reload, resume, and fork all restore the same position; a fork inherits its parent's state through the inherited event prefix.
  • Default: on, matching today's behaviour.

What OFF cuts

ChannelEnforcement
session_search, session_event_search, session_event_read, session_trace, session_event_traceMonotonic tool guard denies the call; a per-agent restriction hides them from the model
cache_search (research page cache in ~/.cache/dsh-free-search/research.db, shared across conversations)same — it can return pages another conversation fetched
Cross-conversation references (@-mentioned session snapshots)filtered out of the prepared step (source.kind = session-reference)
Team relay messages (Agent Teams source.kind = agent-message)filtered by default; blockTeamRelay: false lets them through
Subagent results re-injected into the parent (source.kind = subagent-settled)kept by default (same lineage); blockSubagentSettled: true filters them
Subagent conversations spawned by this oneinherit OFF, including after a restart (the parent's durable projection is consulted when the parent is not live)
Model behavioura memory-switch:policy prompt section explains that memory is off and not to work around it
OpenViking tools, MCP resources, recall, capture, commits, queued writesguard plus wrapped openvikingMemory service (initialize / profileMessage / recallMessage / capture / maybeCommit / flush / dispose). Only when @openviking/dsh-memory-plugin is installed; otherwise this whole section is inert and every other channel still applies

Left on by default (configurable): ThoughtDAG why_* tools (blockThoughtdag: true) and OpenViking-provided skills (blockOpenVikingSkills: true). Workspace AGENTS.md instructions and ordinary skills are not conversation memory and always stay available.

Other MCP memory servers (for example mcp__notion__*) are not blocked by default: whether they count as conversation memory depends on what you wired up, so declare them with denyToolNames / denyToolPrefixes.

Install

A standard DSH profile bundle: host half lib/index.js, browser half lib/client.js (dsh.client). Install it from GitHub:

dsh plugin add github:Kingzhoudk/dsh-memory-switch
# or the explicit git spec
dsh plugin add git+https://github.com/Kingzhoudk/dsh-memory-switch.git

The plugin manager records the dependency and enables the bundle; without HMR the runtime must restart before the Host half mounts. The browser half appears after the page reloads.

For a local checkout, install the directory instead (dsh plugin add link:/absolute/path/to/dsh-memory-switch), or mount it by hand through the profile:

// ~/.dsh/profiles/web/package.json — dependencies
"dsh-memory-switch": "link:/absolute/path/to/dsh-memory-switch"

Mount it either as a bundle (append "dsh-memory-switch" to dsh.profile.bundles; takes effect on restart) or through the profile's user patch layer (patchReload: live, effective immediately):

# ~/.dsh/profiles/web/cordis.patch.yml
- insert:
    - id: memory-switch
      name: dsh-memory-switch

Then run pnpm install inside the profile directory so node_modules/dsh-memory-switch resolves.

Commands

InputBehaviour
/memory or /memory statusshow the current state of this conversation
/memory offcut every cross-conversation memory channel (takes effect immediately)
/memory onrestore every channel
/memory toggleinvert

Configuration

- id: memory-switch
  name: dsh-memory-switch
  config:
    blockOpenViking: true        # OpenViking recall/capture + its tools (needs the plugin)
    blockSessionTools: true      # session_search / cache_search and friends
    blockSessionReferences: true # drop @ cross-conversation snapshots
    blockTeamRelay: true         # drop team relay messages (agent-message)
    blockSubagentSettled: false  # drop subagent result re-injection (subagent-settled)
    blockThoughtdag: false       # also deny why_* (off by default)
    blockOpenVikingSkills: false # also deny OpenViking skills
    denyToolNames: []
    denyToolPrefixes: []

Known limits

  • Direction: an OFF conversation is strongly prevented from reading other conversations, but the reverse — other conversations' searches not finding it — is not achievable in this plugin: the session-query SQLite index and the ThoughtDAG index have no per-session exclusion seam (dsh-thoughtdag reads ~/.dsh/sessions directly). Symmetric isolation needs a change in those packages.
  • Not a sandbox: this plugin only closes model-visible channels. With bash / read available, the model can still read ~/.dsh/sessions/**.jsonl.zstd or ~/.cache/dsh-free-search/research.db directly. Preventing that needs a sandbox policy (denyToolNames: ['bash'], workspace file restrictions), which is out of scope here.
  • Third-party coupling: the OpenViking side wraps the methods of its openvikingMemory service (initialize, profileMessage, recallMessage, capture, maybeCommit, flush, dispose — seven). Missing methods are skipped, but a renamed method after an upgrade would silently weaken that channel (the tool guard is unaffected). The wraps register a ctx.effect restorer, so a live reload does not stack them.
  • When OpenViking is not installed: nothing in ~/.openviking/pending is cleaned (without the service there is no ovSessionId); clear that directory by hand, or reinstall the plugin and toggle once.
  • Already-injected content is not retracted: turning the switch off after a step does not remove recall text already written into this conversation's history; it only stops future reads and writes.

Development

npm test          # same as node --test tests/host.test.mjs tests/client.test.mjs

The browser-half tests resolve React from a DSH source checkout: export DSH_CHECKOUT=/path/to/deepseek-harness to run them; without it those cases are skipped.

node --test tests/ fails on Node 22 with Cannot find module .../tests — Node does not expand a directory argument there, so pass the files.

  • tests/host.test.mjs — argument grammar, projection folding (success / failure / unrelated command), tool guard (including MCP resource tools and cache_search), every cross-conversation message kind, OpenViking wrapping (seven methods), subagent inheritance, pending-queue cleanup including a second OFF, invalid-config fallback and reporting.
  • tests/client.test.mjs — bundle self-registration, slot contract, fail-closed rendering before the projection arrives, and both settled states.

License

MIT — see LICENSE.

相关插件