zilliztech/memsearch#MemSearch

MemSearch plugin for DeepSeek Harness (DSH). It gives DSH persistent, cross-agent memory on the same .memsearch/memory/ markdown store used by the Claude Code, Codex, OpenClaw, and OpenCode plugins, backed by a Milvus hybrid search index.

capture  ── session/event turn/end ──> summarize (dsh-headless agent default, or custom-llm) ──> memory/YYYY-MM-DD.md
inject   ── agent/pre-step step 1  ──> memsearch search ──> returned chunks injected (zero cost otherwise)
recall   ── ctx.skills.register(memory-recall) ──> search → expand → transcript
review   ── web UI dock panel ──> GET/POST /memsearch-dsh/* ──> list candidates / queue review / install

Prerequisites

• Install memsearch:

  uv tool install "memsearch[onnx]"
  

• A DSH profile (web / headless / tui) you want to attach memory to.

• Node >= 22.19 (DSH's requirement).

Install

From npm (recommended)

dsh plugin --profile web add @zilliz/memsearch-dsh

From source (development)

dsh plugin --profile web add /path/to/memsearch/plugins/dsh

dsh plugin is a pnpm forwarder: it links the package into the profile, detects the dsh.bundle declaration in package.json, and appends it to the profile's bundle layers. cordis.patch.yml inside the package then inserts the memsearch row into the profile's plugin tree.

Replace web with your profile name (headless, tui, ...), and restart DSH for the profile (or start a new session) so the plugin mounts.

Manual patch insertion (no dsh plugin)

Append this row to the profile's cordis.patch.yml and make sure @zilliz/memsearch-dsh is resolvable from the profile's node_modules (for example a link: dependency):

- insert:
    - id: memsearch
      name: '@zilliz/memsearch-dsh'

Verify it loaded

Start DSH and check the session log for the plugin mount, or confirm the memory-recall skill is available through the skill tool. Captured turns land in <project>/.memsearch/memory/YYYY-MM-DD.md.

Configuration

The plugin is configured through the profile's cordis.patch.yml config block (patch the memsearch row you inserted). All keys are optional.

Key	Type	Default	Meaning
captureEnabled	bool	true	Capture completed turns into memory.
injectEnabled	bool	true	Inject returned memory candidates before each turn's first step.
summarizeEnabled	bool	true	Summarize turns before writing (on failure a short unavailable note is written, never a raw dump).
summarizeMode	string	auto	Summarizer backend. auto (default) mirrors the other platform plugins: if [plugins.dsh.summarize] provider is set in memsearch config, it uses custom-llm; otherwise dsh-headless (zero-config DSH agent). Explicit dsh-headless / custom-llm pin the backend.

Everything else — provider/model, Milvus, collection, memory dir — comes from memsearch config / environment, exactly like the other platform plugins (no per-plugin config fields):

• Summarize provider/model → [plugins.dsh.summarize] provider / model in ~/.memsearch/config.toml (or [llm.providers.*]; see the custom-llm section below).
• Milvus → [milvus] uri in memsearch config.
• Collection → derived from the project path (derive-collection.sh), or --collection passed to the memsearch CLI.
• Memory dir → MEMSEARCH_DIR env (explicit → global scope), else <project>/.memsearch.

Maintenance tasks (PROJECT.md / USER.md / skills)

Optional background upkeep, aligned with the other platform plugins. Each task is disabled by default; enable the ones you want in ~/.memsearch/config.toml:

[plugins.dsh.project_review]
enabled = true            # maintain .memsearch/PROJECT.md
[plugins.dsh.user_profile]
enabled = true            # maintain .memsearch/USER.md
[plugins.dsh.memory_to_skill]
enabled = true            # distill recurring workflows into skill candidates
min_occurrences = 3       # how often a workflow must recur before distilling

Common settings per task: provider (native = a one-shot DSH headless agent, default), model, min_interval_hours (default 24), input_dir, output_file. Candidates land in .memsearch/skill-candidates/ (git-tracked) and are never installed automatically — installing is a human step (see the memory-to-skill skill in the other platform plugins).

Example override layer (add this to the profile's own cordis.patch.yml):

- id: memsearch
  config:
    summarizeMode: dsh-headless   # pin the headless backend (default is auto)

Summarization modes

Two backends are available, selected by summarizeMode — the same "configured choice" the Claude Code / Codex / OpenClaw / OpenCode plugins offer (each can summarize with their own LLM or a headless agent + small model). The default (auto) matches theirs: configure a provider and you get a direct LLM call; configure nothing and you get a headless agent.

• auto (default) — mirrors the other platform plugins:
  • if [plugins.dsh.summarize] provider is set in memsearch config (~/.memsearch/config.toml, same place the other plugins read), use custom-llm with that provider/model;
  • otherwise use dsh-headless (zero-config DSH agent). This means the plugin behaves like the other four: configure a provider → direct LLM; configure nothing → headless.
• dsh-headless — boots a one-shot DSH headless agent (dsh --profile headless "<summarize task>") to write the notes, mirroring how the other plugins reuse their own agent's headless mode. Zero-config for anyone already using DSH: the sub-agent's model is the deployment's agent-default-model — the user layer of ~/.dsh/settings.yaml (the same selection the Web UI model settings write) wins over any patch, so the [plugins.dsh.summarize] provider/model do NOT apply here — change the model in DSH settings (agent-default-model: in ~/.dsh/settings.yaml, or the Web UI model picker) instead. The boot is asynchronous and fire-and-forget, so the few seconds of headless startup never block the conversation. Requires dsh on PATH or DSH_CLI set to the CLI entry. The sub-agent is booted with MEMSEARCH_DSH_SUMMARIZE=1; the plugin checks that flag and stays inert (no capture / inject / skill) inside the summarizer, so the summarizer's own session is never re-captured in a loop.
• custom-llm — scripts/summarize.py imports memsearch's [llm.providers.*] config and calls the LLM directly. Lightweight: one python process, no DSH boot, no extra CLI dependency. Choose this when you want a specific small model (e.g. an official deepseek-v4-flash key in memsearch config) without booting an agent. Provider selection (most specific first):
  1. [plugins.dsh.summarize] provider (or the summarizeProvider CLI argument summarize.py receives from it) — looked up in [llm.providers.<name>]; a missing entry fails loudly (visible error), never a silent empty write.
  2. llm.provider when it names a configured provider or is a raw type.
  3. compact.llm_provider (deprecated) or openai as a final default.

There is no automatic fallback between modes: the backend you configure (or auto resolves) is the backend used. If it fails (missing dsh CLI, bad provider config), a short unavailable note is written with the reason — the plugin never silently switches to an LLM you did not configure.

A failed summarization writes a short unavailable note (mirroring Claude Code's behavior — memory stays clean, the transcript anchor keeps the raw content reachable for progressive disclosure), and logs a visible warning through the DSH logger.

How it works

• Capture — listens on session/event for turn/end, renders the turn ([User] / [Assistant] / [Tool call] lines), then fire-and-forget summarizes (if enabled) and appends it to the session's own memory/YYYY-MM-DD.md with the shared anchor format <!-- session:<id> turn:<N> db:<path> -->. The project directory comes from session.header.cwd, so a long-lived web surface captures every project it hosts, not just the process's boot directory. Turns are serialized (LLM summarize calls never overlap) and captureExists dedup keeps each turn idempotent if its event replays.
• Inject — on agent/pre-step at step 1, runs a bounded memsearch search over the user's question. Only when the search returns chunks does it inject them plus a [memsearch] Retrieved memory context attached. marker; otherwise the decision is returned unchanged (zero context cost). The model still evaluates whether each candidate chunk is relevant.
• Recall — registers a memory-recall skill (invocable through DSH's native skill tool) that performs search → expand → transcript drill-down and returns a curated summary.
• Maintenance — runs the shared maintenance runner (PROJECT.md / USER.md upkeep and memory-to-skill distillation), triggered on session/disposed plus a 6-hourly fallback timer. Each task is a due-state machine: it runs at most once per min_interval_hours (default 24h) and only when enabled in memsearch config ([plugins.dsh.project_review], [plugins.dsh.user_profile], [plugins.dsh.memory_to_skill]), mirroring the other platform plugins. The maintenance work is executed by a one-shot DSH headless agent (the same dsh --profile headless mechanism as summarization), booted with MEMSEARCH_DSH_SUMMARIZE=1 so the plugin stays inert inside it.
• Skill review panel (web only) — a non-blocking dock strip above the composer (registered into the conversation.input.dock slot) lists skill candidates distilled into .memsearch/skill-candidates/. It is served by the plugin's browser half (client.js, declared through dsh.client in package.json) and talks to the host through two JSON routes on the DSH web server:
  • GET /memsearch-dsh/skill-candidates?sessionId=<id> — lists candidates (pending first) parsed from each candidate's meta.json.
  • POST /memsearch-dsh/skill-review with { sessionId, name, action }:
    • action: "review" queues a [memsearch] Skill candidate ... user message into the live agent's inbox (agent.inbox.append('next-turn', ...)). The agent reviews the candidate on its next turn — non-blocking, it never interrupts a running turn and never opens a blocking dialog.
    • action: "install" runs memsearch skills install <name> --path <dir> in the background (detached, unref'd) to the resolved target: the first entry of plugins.dsh.memory_to_skill.paths in memsearch config (relative entries resolve against the project dir), else the DSH default ~/.agents/skills, which the skill-filesystem provider watches and loads automatically. The project directory is resolved from the session id (its durable cwd), so the panel reflects the project of the session you are viewing on a multi-project web surface. The routes are registered once the DSH web server service becomes available (the memsearch plugin mounts in the base bundle layer, before the web server starts; apply retries on a 1s unref'd timer). Headless / TUI profiles have no browser: the webServer service never appears, no routes are registered, and everything else is unchanged.

Uninstall

dsh plugin --profile web remove @zilliz/memsearch-dsh

Removing the dependency drops the profile-layer entry; the memory markdown files and the Milvus index are left untouched.

Development

• The plugin is plain ESM with no build step — dsh plugin add links the checkout directly, so edits are live after a profile reload.
• The browser half (client.js) is a prebuilt client-module bundle: it registers its factory with window.__ModuleLoader__.load({ id, factory }) (the lazy CJS table the web shell serves at /plugins/<id>/client.js) and exports the Cordis client plugin shape (inject + apply). It is checked in as-is; edit it directly and keep the __ModuleLoader__ registration format (mirror the shipped @deepseek-ai/dsh-client-ui-* lib/client.js bundles).
• Python helpers under scripts/ are linted with the repo's ruff config and tested under plugins/dsh/tests/.
• Keep the memory-write format byte-compatible with the other platform plugins; see plugins/opencode/scripts/capture-daemon.py for the canonical writer.