memweave

memweave command [options] [args]

memweave is a Python library and CLI that gives AI agents a persistent, searchable memory whose primary storage format is plain Markdown. Every memory file is hashed, chunked, and embedded into a single local SQLite database that combines FTS5 keyword ranking with sqlite-vec vector search, so retrieval works offline and merges keyword and semantic signals in one ranked list.Each subcommand maps directly onto a method of the underlying MemWeave Python class, which makes it natural to compose memweave with shell pipelines, editor hooks, and CI jobs without writing Python. Embeddings are cached by content hash, so re-running memweave index is cheap when most files are unchanged, and memweave search never invokes an LLM — only the embedding endpoint.

index

Walk the workspace and embed any Markdown files whose SHA-256 has changed since the last run.

Index a single Markdown file immediately.

Run a hybrid (BM25 keyword + vector) search across the index.

List every tracked file with source labels and chunk counts.

Print a summary of index state, search mode, cache usage, and staleness warnings.

Workspace directory to operate on (default $PWD).

Override the embedding model (e.g. text-embedding-3-small).

Skip change detection and reprocess every file.

Maximum number of search hits to return.

Drop hits below a relevance score.

Restrict results to a labeled source (e.g. sessions).

Search strategy: hybrid (default), keyword, or semantic.

Maximal Marginal Relevance trade-off between relevance and diversity.

Apply temporal decay so older notes rank lower over time.

Emit JSON output suitable for piping into other tools.

memweave stores its database alongside the workspace by default. The active embedding model is selected with --embedding-model or read from environment variables exposed by the chosen provider (e.g. OPENAI_API_KEY for OpenAI embeddings). For the optional file-watcher integration install with the watch extra:

$ pip install memweave[watch]

memweave is intentionally zero-infrastructure — there is no server component, sharing model, or access-control layer. The hybrid index assumes Markdown input; binary files are ignored. Because embeddings are cached by content hash, switching embedding models requires --force to invalidate prior vectors. There are unrelated projects with similar names; this page documents sachinsharma9780/memweave.

memweave was created by Sachin Sharma (sachinsharma9780) and published on PyPI as memweave and on GitHub at github.com/sachinsharma9780/memweave. It is positioned as a lightweight alternative to vector-database stacks for agentic workflows that already use Markdown as a writing format.

sqlite3(1), pip(1), grep(1), ripgrep(1)