Most agent projects struggle with two linked problems: memories are either locked in a single agent provider or stored in opaque blobs that are hard to audit and edit. memsearch takes the opposite approach: Markdown is the source of truth and a vector index is a rebuildable cache. That design makes agent history portable, human-readable, and easy to version-control — which materially lowers friction when you want conversational context to follow users across different agent frontends.
What Sets It Apart
- Markdown-first memory: conversations and summaries are written to daily .md files; the system re-chunks and hashes content so unchanged text isn’t re-embedded. This makes history auditable and editable by humans (so what: easier debugging, compliance, and manual corrections).
- Hybrid, progressive retrieval: a three-layer recall pipeline (search → expand → transcript) combines dense vectors, BM25 sparse retrieval, and RRF reranking to prioritize relevance while allowing cheap, fast fallbacks (so what: better recall for short facts and long transcripts without excessive compute).
- Shadow-index with Milvus: Milvus is used as a rebuildable “shadow” index — the markdown repo remains authoritative and the vector DB is always derivable (so what: you can reset or migrate indexes without losing source data).
- Plugins + cross-agent sync: first-class plugins for Claude Code, OpenClaw, OpenCode, and Codex let conversations captured in one agent become searchable context in others, with zero extra setup for end users.
Who It's For and Tradeoffs
Great fit if you build or operate chat/assistant agents and want persistent, inspectable memory that travels across platforms — especially small teams or developers who prefer local-first options (default ONNX embedding runs locally; model ~558 MB). Look elsewhere if you need a turnkey, multi-tenant enterprise memory service with guaranteed SLA and large-scale managed indexing out of the box; memsearch favors portability, developer control, and a rebuildable architecture over opaque managed abstractions.
Where It Fits
Positioned between simple file-based note-taking and heavyweight managed memory services: use memsearch when you want an auditable, developer-friendly memory layer for RAG/agent workflows, with optional upgrade paths to Zilliz Cloud or self-hosted Milvus for production scaling.
How It Works (brief)
Plugins append summarized turns to Markdown; a file watcher re-chunks changed files, computes SHA-256 hashes, and only re-embeds changed chunks. Embeddings default to local ONNX (bge-m3) but can be switched to OpenAI/Ollama. Upserts go to Milvus (Milvus Lite by default), and retrieval uses hybrid ranking across sparse and dense signals with progressive expansion into full transcripts.
