memsearch CC Plugin — Summary

memsearch (zilliztech/memsearch) is a cross-platform semantic memory system for AI coding agents, built by Zilliz (the company behind Milvus vector database). The Claude Code plugin (plugins/claude-code/) is a lightweight wrapper around the memsearch Python CLI: 4 hooks (SessionStart, UserPromptSubmit, Stop, SessionEnd) that automatically capture conversation turns, a single memory-recall skill (context: fork), and shell scripts that call memsearch CLI commands. Memories are stored as daily .md files in .memsearch/memory/ using Milvus as the vector index (ONNX bge-m3 int8 embeddings by default — no API key required). The system uses 3-layer progressive disclosure: search → expand → transcript. One key design choice: the plugin has NO MCP server — it relies entirely on Claude Code's native hooks + skills + shell scripts.

Differs from seeds: closest to ccmemory (persistent memory via Claude Code plugin, graph + vector search) but memsearch uses Markdown as the source of truth (not Neo4j), Milvus as the shadow vector index (not the primary store), and ONNX local embeddings (not cloud API). Unlike ccmemory's MCP-anchored architecture, memsearch-cc is hooks-and-skills-only with no MCP server.

memsearch CC Plugin — Origin and Philosophy

Origin

Organization: zilliztech (Zilliz, the Milvus vector DB company). 1,841 stars. Apache-2.0 license? (README says MIT badge but org is commercial). Created 2026, active. 14 contributors.

Philosophy: Markdown is the Source of Truth

From the Claude Code plugin README:

"Markdown is the source of truth — inspired by OpenClaw. Your memories are just .md files — human-readable, editable, version-controllable. Milvus is a 'shadow index': a derived, rebuildable cache."

"Progressive retrieval, hybrid search, smart dedup, live sync — 3-layer recall (search → expand → transcript); dense vector + BM25 sparse + RRF reranking; SHA-256 content hashing skips unchanged content; file watcher auto-indexes in real time."

Key Design Decisions

1. No API key required: default embedding is ONNX bge-m3 int8 — runs locally with no GPU. Quality comparable to OpenAI text-embedding-3-small (~1% lower on benchmark).

2. Milvus as shadow index: the Markdown files are the durable store; Milvus is a rebuildable cache derived from them. If Milvus data is lost, memsearch index --force regenerates from .md files.

3. Plugin has NO MCP server: unlike most memory frameworks that add MCP servers, memsearch-cc is hooks + skills + shell scripts only. This avoids sidecar service complexity.

4. Cross-platform sharing: memories captured in Claude Code are queryable from Codex, OpenClaw, OpenCode via the same memsearch CLI/server. One memory store, multiple harnesses.

5. Auto-recall via skill autonomy: memory-recall skill uses context: fork — runs in a forked subagent context so recall doesn't pollute the main agent's context window.

memsearch CC Plugin — Architecture

Distribution

# Claude Code marketplace
/plugin marketplace add zilliztech/memsearch
/plugin install memsearch
# Restart Claude Code

Or standalone Python:

uv tool install memsearch
# OR
pip install memsearch

Required Runtime

• Python 3.10+
• ChromaDB (default vector backend) or Milvus
• ONNX runtime (for local embeddings — auto-downloaded ~300MB)
• memsearch CLI binary

Directory Structure

plugins/claude-code/
├── .claude-plugin/
│   └── plugin.json          # memsearch version 0.4.4
├── hooks/
│   ├── hooks.json           # SessionStart, UserPromptSubmit, Stop, SessionEnd
│   ├── common.sh
│   ├── session-start.sh
│   ├── user-prompt-submit.sh
│   ├── stop.sh
│   ├── session-end.sh
│   └── parse-transcript.sh
├── skills/
│   └── memory-recall/       # Single skill: memory-recall (context: fork)
├── prompts/
├── scripts/
└── transcript.py

Data Flow

Claude Code conversation turn
      ↓ UserPromptSubmit hook
shell script → memsearch CLI → chunking → ONNX embeddings → Milvus index
                                                           → .memsearch/memory/YYYY-MM-DD.md

Query time:
memory-recall skill (forked subagent)
      ↓ memsearch search <query>
3-layer recall:
  1. search: dense vector + BM25 sparse + RRF rerank
  2. expand: related chunks via graph traversal
  3. transcript: original session transcript lookup
      ↓
Context injected into main session

Memory File Location

.memsearch/memory/          # Daily .md files with summaries
.memsearch/memory/YYYY-MM-DD.md

Verify working: ls .memsearch/memory/

Plugin Architecture Summary

memsearch Python library (core: chunker, embeddings, vector store)
      ↓
memsearch CLI (search, index, watch, expand, transcript, config)
      ↓
plugins/claude-code (hooks + skill — wraps CLI in shell scripts)

memsearch CC Plugin — Components

Hooks (plugins/claude-code/hooks/hooks.json)

Skills

Two trigger modes:

1. Explicit: /memory-recall what did we discuss about Redis?
2. Autonomous: Claude auto-invokes when it senses the question needs history

Shell Scripts

memsearch CLI Commands (Python library)

Plugin Marketplace Entry

{
  "name": "memsearch",
  "description": "Automatic semantic memory for Claude Code — remembers what you worked on across sessions",
  "version": "0.4.4",
  "source": "./plugins/claude-code",
  "category": "productivity"
}

memsearch CC Plugin — Uniqueness and Positioning

Differs from Seeds

Closest to ccmemory (persistent memory Claude Code plugin with vector search) but differs in three fundamental ways:

1. Markdown-first vs Neo4j-first: ccmemory uses Neo4j as the primary store. memsearch uses Markdown .md files as primary storage with Milvus/ChromaDB as a rebuildable shadow index. Loss of the vector index doesn't lose the memory — just memsearch index --force to rebuild.

2. No MCP server in the plugin: ccmemory is Archetype 3 (MCP-anchored toolserver). memsearch-cc is hooks + skills + shell scripts only — no sidecar MCP server process required for Claude Code users.

3. Zero external dependencies for default path: ccmemory requires Neo4j + cloud embeddings (or Ollama). memsearch-cc works with a local ONNX model download (~300MB) and ChromaDB (bundled). No cloud API key required.

Also compared to agentmemory (same batch): agentmemory is a Node.js server with 12 hooks and 53 MCP tools; memsearch-cc is a Python plugin with 4 hooks and 1 skill. memsearch-cc is architecturally simpler.

Positioning

memsearch targets developers who want cross-platform semantic memory without infrastructure overhead. The Zilliz backing means enterprise-grade Milvus support is available for users who want cloud-scale vector search, while the default path (ChromaDB + ONNX) needs nothing external. The "Markdown is the source of truth" philosophy appeals to developers who distrust black-box vector stores.

Observable Failure Modes

1. ONNX model download: first-run ~300MB download can fail in offline/restricted environments.
2. Milvus coupling: changing embedding models requires memsearch index --force re-indexing — expensive for large memory stores.
3. Stop hook async timeout: 120s timeout for transcript mining may be insufficient for very long sessions with large transcripts.
4. Cross-session memory quality: automatic summarization quality depends on the embedding model; irrelevant or verbose sessions can pollute the memory store.
5. Plugin-only path has no MCP: Cursor/Windsurf users must use the MCP server path, not the plugin — two separate installation paths.