MemWeave is an open-source Python library that gives AI agents memory a human can actually read. Instead of a vector database, memories are stored as Markdown files on your own disk. That means you can open them in an editor, grep them, and use git diff to see exactly what your agent has learned. Search runs entirely on SQLite, combining BM25 keyword search with vector search in a hybrid pipeline. There is no vector database to provision and no cloud bill. The project is developed by Sachin Sharma, based in Germany. Version 0.1.0 was released on April 3, 2026, and as of August 2026 the latest version is 0.2.1 (released in May 2026). It is licensed under MIT.
Key Features
- Memories are plain Markdown files: Memories live as
.mdfiles under amemory/directory. Subdirectories act as per-agent namespaces, so you can design shared or isolated memory scopes freely. Agent memory, which usually becomes a black box, stays reviewable, editable, and version-controllable as plain text - Hybrid search on SQLite alone: BM25 keyword search via FTS5 and vector similarity search via sqlite-vec run in parallel, and their scores are merged with configurable weights (by default 0.7 for the vector side and 0.3 for BM25). No dedicated vector database server is needed, and the whole index fits in a single SQLite file
- Temporal decay and MMR re-ranking: Scores of older memories decay exponentially based on file date, with a default half-life of 30 days. Files without a date are treated as “evergreen” and are exempt from decay. MMR re-ranking then suppresses near-duplicates so results stay diverse
- No LLM calls during search: Embeddings are cached by content hash, and the search path itself makes zero LLM calls. If the embedding API is unavailable, it falls back to keyword search, so the library keeps working fully offline
- Choose your embedding model: Through LiteLLM it supports OpenAI, Gemini, Mistral, Cohere, and Voyage AI, as well as local models via Ollama. A keyword-only mode that needs no API key is also available
- Both a CLI and a Python API: You can work through command-line operations such as
memweave index/search/files/stats, or through an async Python API likeawait mem.search(...). Background file watching for automatic re-indexing is supported as well - Benchmark results: On the LongMemEval-S long-term memory benchmark, the project reports 98.00% Recall@5 and 99.11% Recall@10, using the local all-MiniLM-L6-v2 embedding model
Pricing
| Plan | Price | What you get |
|---|---|---|
| Open source (MIT) | $0 | Full library and CLI functionality. Install with pip install memweave |
MemWeave itself is free and incurs no server costs. If you use a cloud embedding model such as OpenAI or Gemini, that provider’s API usage fees apply separately. Running a local model through Ollama removes that cost too.
Pricing information is current as of August 2026. Please check the official repository for the latest details.
Pros & Cons
✅ Pros
- Memories are plain Markdown, so you can inspect them by eye, fix them by hand, and track their history in git
- No vector database to build or operate. Everything fits in a single SQLite file, keeping adoption cost low for personal projects and small teams
- Combining keyword and semantic search handles both exact proper nouns and vague phrasing well
- Temporal decay structurally mitigates the problem of stale information crowding out newer information
- Fully offline operation is possible, so sensitive data never has to leave your machine
- The MIT license imposes few restrictions on commercial use
⚠️ Cons
- Python 3.12 or later is required, which is a relatively new baseline
- It is a library rather than a finished SaaS, so there is no UI or admin console, and integration requires implementation work
- At version 0.2.1 the project is still early, and APIs and configuration options may change
- As a solo-developer project, enterprise-grade support or an SLA should not be expected
- For workloads with large-scale, high-frequency writes, the SQLite-and-files design may become a constraint
Comparison with Similar Services
| Criteria | MemWeave | Mem0 | Zep | Letta |
|---|---|---|---|---|
| Memory storage format | Markdown files + SQLite | Proprietary store (vector/graph) | Temporal knowledge graph | Agent state managed in a database |
| Human-readable directly | Readable and hand-editable | Mainly via API | Mainly via API | Mainly via API |
| Infrastructure required | None (one SQLite file) | Self-hosted or cloud | Self-hosted or cloud | Self-hosted or cloud |
| Delivery model | Python library + CLI | Library + managed service | Primarily a managed service | Agent platform |
| Main scope | Focused on storing and searching memory | General memory layer | Conversation history and relationship tracking | Building stateful agents |
MemWeave positions itself narrowly around storing and searching memory, which is a different role from Letta, a platform for building agents themselves. Compared with Mem0 and Zep, its differentiators are that memories remain human-readable files and that it requires no additional infrastructure at all.
Who Is It For
- Individual developers who want long-term memory for AI agents but would rather avoid operating a vector database
- Anyone who wants to be able to review and correct what an agent has remembered
- Teams that need to implement memory locally without sending confidential data to external services
- People who want to turn existing Markdown notes and documentation into an agent’s knowledge source as-is
- Development teams that want memory content managed in git so changes can be reviewed
Summary
MemWeave is built on the idea that an agent’s memory should live in readable files rather than an opaque vector space. It delivers hybrid BM25 and vector search using nothing but SQLite, keeps practical accuracy through temporal decay and MMR, and holds infrastructure cost at zero. It is still a young project in the 0.2 series, so if you build it into a production system, leave room to follow API changes. For personal projects and internal tools where you simply want to give an agent memory and see how it goes, its light footprint and transparency are significant advantages.