What is Memvid?
Memvid stores an agent's memory in a single file. A .mv2 file holds a header, an embedded write-ahead log for crash recovery, compressed data segments, a Tantivy BM25 full-text index, an HNSW vector index and a time index, with no sidecar files. Writes are append-only Smart Frames, each with a timestamp and checksum, so you can query or replay an earlier state of the memory.
The name is a leftover. Version 1, written in Python, stored memories as QR codes inside video frames; the January 2026 rewrite in Rust (v2.0.0) dropped that for a binary format, and the project now asks that the QR-code description not be repeated. Optional features add PDF extraction, CLIP image search, Whisper transcription, OpenAI embeddings and AES-256-GCM encryption.
Memvid is a library, not a fact-extraction service: it indexes what you put in it. The core includes a regex rules engine that pulls entity, slot and value memory cards from text without a model, and there are Python and Node.js SDKs and a CLI. It is Apache 2.0 licensed.
Is Memvid worth it?
Memvid fits a narrower job than Hindsight or Zep. It suits agents that must run offline or on a laptop, ship their knowledge as an artefact, or prove exactly what they knew at a given time: one .mv2 file holds the text, a BM25 index, an HNSW vector index and a timeline, and can be copied, encrypted or rewound. Nothing has to be running.
What it does not do is decide what is true now. Writes are append-only frames; the rules engine extracts memory cards with regular expressions, and the docs do not describe how a contradicted fact is superseded. A single writer holds an exclusive lock, so concurrent agents sharing one file queue behind each other. For a user whose details change, you are left ranking old and new frames yourself, which is the problem Graphiti and Hindsight are built to solve.
The benchmark claim needs context. Memvid reports 85.65% on LoCoMo categories 1 to 4, with the adversarial category excluded, GPT-4o answering and GPT-4o-mini judging, and calls this 35% better than rivals. The rival figures, such as Mem0 at 66.88% and Zep at 65.99%, come from Mem0's April 2025 paper and were not rerun; Mem0 now reports 92.5 on LoCoMo for its own platform. The README's latency of 0.025 ms at p50 and 0.075 ms at p99 comes with no corpus size or hardware.
The open-source core has a commercial meter in it. The source sets a default capacity of 50 MB per file, and lifting it to 2 GB or 10 GB requires an Ed25519-signed ticket issued by memvid.com; a write past the limit fails with CapacityExceeded. Apache 2.0 lets you change that, but out of the box a file that holds embeddings and indexes as well as text reaches 50 MB sooner than its raw text suggests.
Development has slowed. The last release was v2.0.140 on 27 May 2026, which fixed write-ahead log corruption after file growth, and the only commits since were README edits in July. The repo has about 17,400 GitHub stars.
Where it earns its place
A field-service company equipping 200 technicians with an offline repair assistant
Search manuals and past repair notes on a tablet with no signal
Build a .mv2 file from PDFs with the pdf_extract and vec features, encrypt it, and sync it to each device; local BGE-small embeddings mean no API at query time. Keep each file under the 50 MB default or obtain a larger-capacity ticket.
A compliance engineer at a fintech running an AI assistant over policy documents
Show auditors exactly what the assistant knew on a given date
Append policy versions as frames and use the time index to replay the state of memory on any date. The file's Blake3 checksums and signatures help show it was not altered.
A Rust developer building a desktop note-taking app
Add hybrid keyword and semantic search without bundling a database
Add memvid-core with the lex and vec features and write each note with put; one file per user keeps backups to a single copy. No LLM is needed unless you choose OpenAI embeddings.
Before you commit
- Setup
- Add memvid-core 2.0 to Cargo.toml with the features you need (Rust 1.85.0+), or install the Python or Node.js SDK; local text embeddings need the ONNX model files downloaded manually.
- Integrations
- Rust core with Python and Node.js SDKs and a CLI; integrations listed for LangChain, LlamaIndex, Vercel AI SDK, OpenAI, Google ADK, AutoGen, CrewAI, Semantic Kernel and Haystack.
- Data and privacy
- Everything stays in the local .mv2 file, with optional password-based AES-256-GCM encryption; nothing leaves the machine unless you enable OpenAI embeddings. No telemetry code appears in the core library. Capacity tickets are issued by memvid.com.
- Cost at scale
- No licence fee and no server; the cost is disk and, if used, embedding calls. Files beyond 50 MB need a capacity ticket from memvid.com, whose prices are not published.
- Limits
- Single writer per file; the write-ahead log scales from 1 MB to 64 MB with file size; no fact update or contradiction handling is documented; last release May 2026.
What’s good
- One portable file with text, BM25, vector and time indexes
- No server, database or network needed at query time
- Append-only frames allow replaying earlier memory states
- Optional AES-256-GCM encryption and PII masking
What’s not
- No documented handling of contradicted or changed facts
- 50 MB default cap lifted only by a signed ticket from memvid.com
- Single writer per file
- LoCoMo comparison uses other vendors' 2025 figures
Best for
Developers building offline or edge agents that need a searchable, portable knowledge file rather than a model of a changing user.
Memvid alternatives
Memvid: common questions
- Is Memvid worth it?
- Memvid is a tidy way to ship searchable memory as one file an agent can carry offline, but it is a document index rather than a memory layer that understands changing facts, and its benchmark lead rests on rivals' 2025 figures. We rate it 3.6 out of 5.
- How much does Memvid cost?
- Free under Apache 2.0. The library caps a file at 50 MB by default (the Free tier in the source); larger capacities of 2 GB and 10 GB are unlocked by signed tickets from memvid.com. Cloud plan prices are not published.
- What are the downsides of Memvid?
- No documented handling of contradicted or changed facts; 50 MB default cap lifted only by a signed ticket from memvid.com; single writer per file; loCoMo comparison uses other vendors' 2025 figures.
- Who is Memvid best for?
- Developers building offline or edge agents that need a searchable, portable knowledge file rather than a model of a changing user.
- How much does Memvid cost for a team or in production?
- No licence fee and no server; the cost is disk and, if used, embedding calls. Files beyond 50 MB need a capacity ticket from memvid.com, whose prices are not published.
- Is Memvid safe for company data?
- Everything stays in the local .mv2 file, with optional password-based AES-256-GCM encryption; nothing leaves the machine unless you enable OpenAI embeddings. No telemetry code appears in the core library. Capacity tickets are issued by memvid.com.
- What are the best Memvid alternatives?
- The closest alternatives we have reviewed are Hindsight, Mem0, claude-mem and Cognee.
Answers checked Nov 6, 2026.