Skip to content
dsh marketplace

xuanling-dsh-memory

Verified

@xuanling-rs/xuanling-dsh-memory · v0.10.0 · MIT

Recommended DeepSeek Harness bundle: add the complete XuanLing Memory v2 profile while retaining DSH's native tools.

Install

dsh plugin add @xuanling-rs/xuanling-dsh-memory

Confirm the layer applied with dsh --profile default --dump-config — see the install guide.

Source

Published to npm without a public repository. Inspect the package contents before installing.

Readme

xuanling-dsh-memory

English | Simplified Chinese

Recommended DeepSeek Harness bundle for XuanLing Memory v2: nine proposal-first memory tools, added next to DSH's native tools. Package @xuanling-rs/[email protected].

What it is

Adds XuanLing's complete Memory v2 lifecycle to a DeepSeek Harness (DSH) profile: memory_search and memory_get for recall, five memory_candidate_* tools for pending proposals, memory_review for the explicit approve/reject decision, and memory_feedback. Writes stay pending until a review activates or rejects them, and the server never claims that a human performed that review. Native DSH tools stay enabled; no filesystem, process, artifact, or session tools are added.

Why use it

  • Use it for durable, cross-project recall that cannot silently rewrite canonical memory, with the smallest recommended tool surface.
  • Use xuanling-dsh-tools when you also need XuanLing filesystem, process, project, artifact, or session tools.
  • Use xuanling-dsh-tools-replace only for XuanLing semantics behind DSH's native text file tools.
  • Pair it with xuanling-dsh-skills for the proposal-first workflow and a host approval gate. Install exactly one XuanLing runtime bundle per profile; see the bundle guide.

Requirements

  • DSH with the shipped web or headless profile.
  • Node.js 22.14 or newer.
  • A DSH build that provides @deepseek-ai/dsh-mcp-client from the 0.2.1-rc.2 capability channel.
  • macOS ARM64, Linux x64 with glibc 2.35+, or Windows x64 with MSVC.

Install

dsh plugin --profile web add @xuanling-rs/[email protected]

Use --profile headless for the headless profile. A custom new profile name starts from the base bundle only and is not a runnable Web or Headless app. The command installs @xuanling-rs/[email protected] inside the profile; no global npm package or npx is needed.

Quickstart

  1. Verify the package, then restart the profile and refresh DSH:

    dsh plugin --profile web list
dsh --profile web --dump-config

  2. Ask the model to search mcp_catalog__xuanling for memory_search and activate that exact raw name. It appears as mcp__xuanling__memory_search on the next request.

  3. Call it with a narrow read-only query. Do not create or review a candidate just to verify an install.

Configuration

  • No environment variable is required; this bundle adds no filesystem tools and does not use XUANLING_WORKSPACE_ROOT.
  • Store: the shared default ~/.xuanling/memory.db, opened lazily on the first valid Memory call. Startup and catalog discovery do not create it.
  • Scopes are global, project (project_id), or workspace (project_id + workspace_id); ancestor scopes are walked only when asked.
  • memory_review requires an explicit user decision naming the proposal; an unreviewed candidate is a valid end state.
  • Model-facing flow: only mcp_catalog__xuanling is visible initially; search it and activate one exact raw name per call.
  • Tool-call budget: 120 seconds. Public tool names are mcp__xuanling__*.

Limits and safety

  • Memory only. No file or process tools, and no filesystem capability.
  • Shared store. ~/.xuanling/memory.db is used by other XuanLing deployments too; never put secrets or private user content in it.
  • Proposal-first is a state machine, not a human attestation. reviewer_id is caller-attested; the Skills bundle adds the DSH approval gate.
  • No startup side effects. The database is neither created nor migrated until the first valid Memory call.

Troubleshooting

  • Only mcp_catalog__xuanling is visible. Expected. Search it and activate exact raw names, one per call.
  • An activated tool does not appear. Use the raw name without the mcp__xuanling__ prefix; activations do not survive a restart.
  • A candidate never becomes active. It is awaiting review; call memory_review only on an explicit user decision.
  • Startup fails. Check Node.js 22.14+, platform support, and reinstall the exact 0.10.0 version. Remove any other XuanLing runtime bundle.

Links