MemSearch is a persistent, cross-platform memory layer for AI coding agents. Plugins for Claude Code, Codex, DeepSeek Harness, OpenClaw, and OpenCode capture each conversation turn automatically, summarize it, and append it to a daily Markdown journal. A conversation held in one agent becomes searchable context in all the others, with no commands to learn and no manual saving.
Markdown files are the source of truth, an approach adopted from OpenClaw: memories are plain files that a person can read, edit, and version-control. Milvus is a shadow index, a derived and rebuildable cache that serves retrieval. A file watcher re-chunks a changed file, hashes each chunk with SHA-256, skips unchanged chunks without an API call, and embeds and upserts the rest. Milvus Lite (a single local file) is the default backend, with Zilliz Cloud or a self-hosted Milvus server as alternatives for teams.
Recall is a three-layer progressive search: hybrid search (dense vectors plus BM25 sparse vectors fused with reciprocal rank fusion) returns ranked chunks, a chunk can be expanded to its full Markdown section, and the original transcript can be parsed for the raw dialogue. Optional remote reranking with Jev reorders results without a local model download. Embeddings default to a local ONNX bge-m3 model on CPU, with OpenAI, Google, Voyage, Jina, Mistral, and Ollama as alternatives.
Each plugin keeps its platform's native summarizer by default (Claude Code uses a Haiku model, for example), and summarization can be routed through a MemSearch-managed API provider instead. Two optional background layers sit on top of the episodic journals: a maintained PROJECT.md (active threads, decisions, risks, next steps) and USER.md (preferences and working style), both off by default, and Skills from Memory, which distills repeated workflows into installable skills following the Agent Skills standard, stored as git-tracked candidates that stay inert until installed.
For agent developers, the same engine is available as a CLI and a Python API (index, search, expand, watch, compact) for adding memory to a custom agent or a memory-augmented RAG pipeline, with examples for OpenAI, Anthropic, and a fully local Ollama setup.
Features
- Agent plugins: Claude Code, Codex, DeepSeek Harness, OpenClaw, and OpenCode plugins that capture every conversation turn and recall relevant context automatically
- One memory across agents: a conversation in one agent becomes searchable context in all the others
- Markdown source of truth: memories are human-readable, editable, version-controllable files, and Milvus is a rebuildable shadow index
- Hybrid search: dense vectors plus BM25 sparse vectors with reciprocal rank fusion, and optional remote reranking with Jev
- Progressive recall: ranked chunks, expansion to the full Markdown section, then the raw transcript when needed
- Live sync and dedup: a file watcher re-indexes changed files, and SHA-256 chunk hashing skips unchanged content
- Embedding providers: local ONNX bge-m3 by default, or OpenAI, Google, Voyage, Jina, Mistral, and Ollama
- Milvus backends: Milvus Lite as a single local file, Zilliz Cloud, or a self-hosted Milvus server
- Summarization routing: each plugin's native summarizer by default, or a MemSearch-managed API provider per plugin
- Background maintenance: optional PROJECT.md and USER.md notes refreshed after sessions when journals change
- Skills from memory: repeated workflows distilled into installable Agent Skills, kept as git-tracked candidates until a person installs one
- CLI and Python API: index, search, expand, watch, and compact for building memory into custom agents and RAG pipelines
