Memvid is not another vector database. It’s a portable, single-file memory system with flexible embedding options. Use local models for privacy and offline use, or connect to external providers like OpenAI, NVIDIA, and more.
The Problem with Traditional RAG
Most memory systems today follow the same pattern:
This approach has serious limitations:
| Problem | Impact |
|---|
| API dependency | Can’t work offline, costs money per query |
| Embedding drift | Model updates break your index |
| No exact matching | ”Error 404” doesn’t find “error 404” |
| Black box relevance | Hard to debug why results are wrong |
| Cold start | Need to embed everything before first search |
The Memvid Innovation
Memvid takes a completely different approach:
Flexible embedding options. Memvid combines multiple search strategies:
- BM25 lexical search - Battle-tested, fast, explainable
- Vector embeddings - Local models (Nomic, BGE, GTE) or external APIs (OpenAI, NVIDIA)
- SimHash deduplication - Find near-duplicates instantly
- Time-aware retrieval - When something was added matters
- Hybrid search - Combines lexical + semantic for best results
How Frames Change Everything
Traditional systems store “chunks” - arbitrary text splits. Memvid stores Frames - structured units of memory:
| Field | Example |
|---|
content | ”The quarterly revenue exceeded $10M…” |
uri | reports/q4-2024.pdf |
timestamp | 2024-12-15T10:30:00Z |
content_hash | blake3(…) |
simhash | 0x8f3a2b1c… |
- What it contains (content + hash)
- Where it came from (URI + metadata)
- When it was created (timestamp)
- What it’s similar to (SimHash)
- What entities it mentions (Logic Mesh)
- How it connects to other frames (relationships)
Frames are the foundation of Memvid’s multi-index approach, enabling both lexical and semantic search.
Multiple Indices, Flexible Providers
Every .mv2 file contains multiple search indices:
| Index | Purpose | How It Works |
|---|
| Lexical (BM25) | Keyword search | TF-IDF scoring, exact matches |
| Vector | Semantic similarity | Local or external embedding models |
| SimHash | Near-duplicate detection | 64-bit locality-sensitive hash |
| Time | Temporal queries | B-tree on timestamps |
| Logic Mesh | Entity-relationship | Triple store (subject-predicate-object) |
- Local models (Nomic, BGE-small, BGE-base, GTE-large) - Fast, private, works offline
- External APIs (OpenAI, NVIDIA) - Higher quality, no local compute needed
Choosing Your Embedding Provider
Memvid supports multiple embedding providers. Choose based on your needs:
| Provider | Models | Best For |
|---|
| Local (Default) | bge-small, bge-base, nomic, gte-large | Privacy, offline use, no API costs |
| OpenAI | openai, openai-small, openai-ada | Highest quality, multilingual |
| NVIDIA | nvidia | Enterprise, high throughput |
# Use local embeddings (default - works offline)
memvid create notes.mv2
memvid put notes.mv2 --input docs/ --embedding
# Use OpenAI embeddings (requires OPENAI_API_KEY)
memvid put notes.mv2 --input docs/ --embedding -m openai
# Use specific local model
memvid put notes.mv2 --input docs/ --embedding -m bge-base
Logic Mesh: Relationships Without ML
Traditional systems need expensive NER models for entity extraction. Memvid’s Logic Mesh uses:
- Rule-based extraction - Fast, free, no API
- Pattern matching - Dates, emails, numbers
- Co-occurrence - Entities mentioned together
- Temporal reasoning - When facts changed
# Enable Logic Mesh during ingestion
memvid put memory.mv2 --input docs/ --logic-mesh
# Query relationships
memvid follow traverse memory.mv2 --start "John" --link "works_at"
# Result: John → works_at → Acme Corp (since 2024-01-15)
SimHash: Smart Deduplication
Instead of comparing embeddings, Memvid uses SimHash - a locality-sensitive hash that detects near-duplicates:
Document A: "The quick brown fox jumps over the lazy dog"
Document B: "The quick brown fox leaps over the lazy dog"
SimHash A: 0x8f3a2b1c4d5e6f70
SimHash B: 0x8f3a2b1c4d5e6f71
Hamming distance: 1 bit → Near duplicate detected!
- Instant - O(1) comparison
- No API calls - Computed locally
- Works offline - No internet needed
- Deterministic - Same input = same hash
Memvid’s local-first approach delivers fast performance:
| Operation | Traditional Vector DB | Memvid (Local Embeddings) |
|---|
| Ingest 1,000 docs | 5-10 minutes | 30 seconds |
| First search | After embedding | Instant |
| Offline search | ❌ | ✅ |
| Cost per query | $0.0001+ | $0 (local) |
| Single file portability | ❌ | ✅ |
Getting Started
Get started with Memvid in minutes:
# Install
npm install -g memvid-cli
# Create memory
memvid create my-memory.mv2
# Add your documents (uses local embeddings by default)
memvid put my-memory.mv2 --input ./documents/
# Search with hybrid lexical + semantic
memvid find my-memory.mv2 --query "quarterly report"
# Ask questions
memvid ask my-memory.mv2 --question "What were the Q4 results?"
Switching Embedding Providers
Need higher quality embeddings? Switch to an external provider:
# Rebuild vector index with OpenAI embeddings
memvid doctor my-memory.mv2 --rebuild-vec-index -m openai
# Search with semantic mode
memvid find my-memory.mv2 --query "financial performance" --mode sem
# Or use hybrid (lexical + semantic)
memvid find my-memory.mv2 --query "financial performance" --mode hybrid
Next Steps
