The memory problem in AI coding agents just got its Markdown solution — and it works across every agent you use.
The Summary
- Zilliz released MemSearch, a cross-platform memory system for AI coding agents that stores conversation history as Markdown files with semantic search built in.
- Works as a plugin for Claude Code, OpenClaw, and other agents — conversations in one agent become searchable context in all others automatically.
- Three-layer hybrid search (dense vector + BM25 sparse + RRF reranking) with real-time file watching and auto-indexing.
- Markdown files are the source of truth; Milvus vector database is just a rebuildable cache.
The Signal
AI coding agents have a memory problem. You spend an hour debugging with Claude Code, figure out your Redis TTL settings, then switch to another agent the next day and it has no idea what you're talking about. Every conversation starts from zero. Every context window is an island.
MemSearch solves this by making memory portable and human-readable. The core insight: store agent conversations as daily Markdown files, then layer semantic search on top. When you ask "what did we discuss about Redis?" across any agent that has the plugin installed, you get answers pulled from every conversation you've ever had, regardless of which agent you were using at the time.
"Markdown is the source of truth — Milvus is a 'shadow index': a derived, rebuildable cache."
The architecture choice matters:
- Your actual memories live in `.memsearch/memory/` as dated `.md` files you can read, edit, grep, or version control
- The Milvus vector database indexes those files for semantic search but can be deleted and rebuilt anytime
- SHA-256 content hashing skips re-indexing unchanged content
- A file watcher auto-indexes new conversations in real time
This inverts the usual vector database pattern. Most agent memory systems treat the database as primary storage and give you an API. MemSearch treats Markdown as primary and the database as infrastructure. If Milvus breaks, your memories are fine. If you want to switch vector databases later, you just reindex the Markdown.
The retrieval is legitimately sophisticated. Three-layer progressive recall starts with semantic search, expands to related context, then pulls full conversation transcripts. Hybrid search combines dense embeddings with BM25 sparse vectors and reciprocal rank fusion reranking. This isn't a toy.
For developers building agents, MemSearch exposes a full Python API and CLI. For users who just want their agents to remember things, it's a one-line plugin install. The plugin auto-captures every conversation turn with zero configuration. You can trigger recall explicitly with `/memory-recall` or just ask naturally and the agent invokes it when needed.
The cross-platform piece is the real unlock. Install the plugin in Claude Code and OpenClaw, and suddenly both agents share the same memory pool. A debugging session in one becomes context for the other. Your agent memory becomes infrastructure that lives above any single agent platform.
The Implication
If you're using AI coding agents regularly, install this. The Markdown-first approach means you're not locked into anyone's proprietary memory format, and the semantic search actually works across sessions.
For agent developers, this is the memory layer you've been duct-taping together yourself. Open source, proven vector search from Zilliz (they make Milvus), and a design philosophy that treats human readability as a feature, not a compromise. Watch how fast this becomes table stakes for any serious coding agent.