Search skills/

memsearch Claude skills

Cross-agent persistent memory for Claude Code, Codex, OpenClaw, OpenCode and DeepSeek Harness: Markdown files as the source of truth, with a Milvus search index.

zilliztech/memsearch 17 skills Memory & context
Our verdict Recommended with caveats

A capable memory layer from a vector database vendor; automatic capture and optional cloud embedding need a deliberate privacy decision.

3/5 Depth expertise shipped ! Security Broad access 2d Maintenance Updated 2 days ago 2.8K Adoption stars OK License MIT, commercial OK

What memsearch does

memsearch gives coding agents a memory that survives across sessions and across tools. Plugins for Claude Code, Codex, OpenClaw, OpenCode and the DeepSeek Harness use hooks to capture each conversation turn and summarize it into Markdown files, which are the source of truth; Milvus is only a rebuildable index. Recall is progressive in three layers: search, expand a hit, then open the transcript. Search combines dense vectors, BM25 and reciprocal-rank reranking, content hashing skips unchanged text, and a file watcher keeps the index current. A conversation in one agent becomes searchable context in the others.

The repo also ships three skills per platform: memory-recall, memory-config (diagnose settings, summarization, index health, provider routing) and memory-to-skill, which distills workflows you repeat into reusable skills, a procedural-memory layer kept current in the background. Optional background maintenance keeps PROJECT.md and USER.md notes updated. Embeddings can run locally with ONNX bge-m3 (about 558 MB downloaded from HuggingFace on first launch), through Ollama, or through OpenAI with a key. Milvus Lite is the zero-config default, with Docker or Zilliz Cloud for teams. A Python API is available for building your own agents.

Good for

  • Keeping context across sessions and between several coding agents
  • People who want memory stored as readable Markdown
  • Developers building memory into their own agents via the Python API

Not for

  • Users who do not want conversations captured automatically
  • Machines that cannot download a 558 MB local model

What's inside depth 3/5

17 skills averaging 7.9 KB with per-platform reference files and a Python package behind them, but no skill-level evals.

Skills
17
Reference docs
60
Scripts
0
Tests / evals
0
Templates
0
Avg SKILL.md
7.7 KB

Counted from the repository file tree. How we score depth.

Dependencies and costs

  • Python package memsearch
  • Milvus Lite (default), or Docker Milvus or Zilliz Cloud with a token
  • Embedding provider: local ONNX bge-m3, Ollama, or OpenAI with OPENAI_API_KEY
  • Telemetry: none found in the files we reviewed; text sent to an embedding or summarizing provider you configure leaves your machine.

Security review

Broad access

This skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.

  • Pre-approves broad shell commandsThe memory-recall, memory-config and memory-to-skill skills (claude-code and opencode copies) declare allowed-tools: Bash with no command pattern, pre-approving unrestricted shell; memory-recall and memory-config also run as a forked context.plugins/claude-code/skills/memory-recall/SKILL.md:5

A cross-agent long-term memory plugin whose skills search, configure and distill your past sessions through the memsearch CLI, with unrestricted Bash pre-approved in several skills. Once installed it captures every conversation turn through hooks and stores summaries locally by default (local Milvus Lite and ONNX embeddings), so recorded conversations may include secrets. Cloud embedding or summarizing providers are opt-in. The hooks and the CLI itself were not in the reviewed copy.

Worth knowing

  • installs_hooksAfter the plugin is installed it captures every conversation turn automatically through documented host hooks (Stop/SessionStart and similar), summarizes it with an LLM and writes it to .memsearch/memory/ for later recall; conversations may contain secrets. Capture can be disabled with memsearch config set plugins.<platform>.summarize.enabled false. The hook definitions and scripts themselves are not in the reviewed copy.README.md:64
  • Runs remote code at an unpinned versionSkills fall back to uvx memsearch / uvx --from memsearch[onnx] memsearch, and memory-config suggests uv tool install -U "memsearch[onnx]", all without a version pin.plugins/claude-code/skills/memory-recall/SKILL.md:21
  • Installs packages or toolsInstalls and upgrades the memsearch Python package with uv; the default local embedding model (ONNX bge-m3, about 558 MB) is downloaded from HuggingFace Hub on first launch.plugins/_shared/skills/memory-config/SKILL.md:67
  • Needs an API key for its own serviceOptional providers are configured with env: references such as env:OPENAI_API_KEY, ANTHROPIC_API_KEY and GEMINI_API_KEY, and a Milvus/Zilliz token; they are meant for the corresponding service only. Text sent to a configured embedding or summarizing provider leaves your machine.plugins/_shared/skills/memory-config/SKILL.md:167
  • Makes network requestsmemory-config curls pypi.org and raw.githubusercontent.com/zilliztech/memsearch (main branch) to compare plugin and CLI versions; storage defaults to a local Milvus Lite file, with Zilliz Cloud or a Milvus server as opt-in remote backends.plugins/claude-code/skills/memory-config/references/claude-code.md:11
  • Changes files outside your projectSetup edits global memsearch config (~/.memsearch), and for OpenClaw runs openclaw config set plugins.entries.memsearch.hooks.allowConversationAccess true and allowPromptInjection true.README.md:151
  • Relies on a CLI or package we could not reviewEvery skill drives the memsearch CLI (search, expand, transcript, config, skills add); that Python package and the plugin hook scripts such as derive-collection.sh are not in the reviewed files.plugins/claude-code/skills/memory-recall/SKILL.md:14

What we checked

The repository (338 files) on Oct 29, 2026. Not a full review: 254 of 338 files were not downloaded (large repository; we reviewed the skill folders); it relies on code outside the repository: memsearch PyPI package and CLI source; Plugin manifests (.claude-plugin/marketplace.json, plugin.json), hook definitions and helper scripts such as scripts/derive-collection.sh (about 254 readable files are not in the local copy); HuggingFace ONNX bge-m3 model, Milvus/Zilliz Cloud services and any embedding or summarization provider you configure.

NVIDIA SkillSpector static scan: 22 pattern hits (MCP Rug Pull 10, Memory Poisoning 6, Supply Chain 6). The 22 hits are 'uvx memsearch' unpinned-package mentions, a 'delete state' phrase in an OpenCode reference page, and a curl of PyPI JSON piped into python3 to print the latest version number.

A review, not a certification. Skills change after we check them — read the files before installing.

License

MIT — commercial use allowed.

Can it run unattended?

Not really. It is a background memory service rather than a task runner, with no task loop or stop conditions.

Alternatives to memsearch

obsidian-wiki

Ar9av/obsidian-wiki

Forty skills that let an agent build and maintain a linked Obsidian vault from your notes, code and past agent sessions, then answer questions with citations.

Recommended Check first Depth 3/5

Context Mode

mksglu/context-mode

MCP server plus hooks that keep bulky tool output out of the context window, index it in local SQLite, and restore session state after compaction.

With caveats Broad access Depth 3/5

memU

NevaMind-AI/memU

Memory layer that mines agent session logs into a shared markdown wiki and re-injects it across Codex, Claude Code, Cursor and others, locally or via memU Cloud.

With caveats Broad access Depth 4/5

Planning with Files

OthmanAdi/planning-with-files

Keeps task_plan.md, findings.md and progress.md on disk and re-injects them via hooks, so a plan survives /clear, compaction and crashes.

Recommended Broad access Depth 4/5

FAQ

What is memsearch?

It is a memory layer for AI coding agents that saves summaries of your conversations as Markdown and indexes them in Milvus. Claude Code, Codex, OpenClaw, OpenCode and DeepSeek Harness plugins can all search the same memory.

How do I install memsearch?

Pick the plugin for your agent in the README, for example the Claude Code plugin, then chat as usual. Python users can install the memsearch package and choose an embedding provider.

Does memsearch need an API key?

Not necessarily. ONNX bge-m3 and Ollama run locally without a key, while OpenAI embeddings need OPENAI_API_KEY and Zilliz Cloud needs a token.

Sources: GitHub repository (stars, license, file tree, last push Oct 27, 2026). Review based on the README and SKILL.md files as published.