Memvid

Memvid is an open-source, portable AI memory system that packages data, embeddings, search structures, and metadata into a single .mv2 file. Instead of running complex RAG pipelines or server-based vector databases, Memvid enables fast retrieval directly from the file, making it model-agnostic and infrastructure-free. It gives AI agents persistent, long-term memory they can carry anywhere, with sub-5ms local access and crash-safe, append-only writes.

• Smart Frames — Memvid organizes memory as an append-only sequence of immutable Smart Frames, each storing content with timestamps, checksums, and metadata for efficient compression and parallel reads.
• Single-File Format (.mv2) — All data, indexes (full-text, vector, time), and metadata live in one portable file with no sidecar files, WAL files, or lock files.
• Time-Travel Debugging — Rewind, replay, or branch any memory state to inspect how knowledge evolves over time.
• Multi-Modal Search — Supports BM25 full-text search, HNSW vector similarity search, CLIP visual embeddings for image search, and Whisper audio transcription.
• Local & Cloud Embeddings — Use local ONNX models (BGE-small, BGE-base, Nomic, GTE-large) or cloud API embeddings via OpenAI, with model-binding to prevent accidental mixing.
• Multi-Language SDKs — Available as a Rust crate (memvid-core), Node.js SDK (@memvid/sdk), Python SDK (memvid-sdk), and a CLI (memvid-cli via npm).
• Offline-First — Works fully offline with no server dependencies, making it ideal for edge deployments, air-gapped systems, and resource-constrained environments.
• Capsule Context — Self-contained, shareable .mv2 memory capsules support rules, expiry, and encryption for secure knowledge sharing.
• Benchmark Performance — Achieves +35% SOTA on LoCoMo long-horizon recall, 0.025ms P50 latency, and 1,372× higher throughput than standard memory systems.
• Broad Use Cases — Supports long-running AI agents, enterprise knowledge bases, codebase understanding, customer support agents, and auditable AI workflows.