memvid-sdk 2.0.112

Single-file AI memory system for Python. Store, search, and query documents with built-in RAG.

memvid-sdk

Single-file AI memory system for Python. Store documents, search with BM25 ranking, and run retrieval-augmented generation (RAG) queries from a portable .mv2 file.

Built on Rust with PyO3 bindings. No database setup, no network dependencies, no configuration files.

Install

pip install memvid-sdk

Optional extras for framework integrations:

pip install "memvid-sdk[langchain]"     # LangChain tools
pip install "memvid-sdk[llamaindex]"    # LlamaIndex query engine
pip install "memvid-sdk[openai]"        # OpenAI function schemas
pip install "memvid-sdk[crewai]"        # CrewAI tools
pip install "memvid-sdk[full]"          # All integrations

Quick start

from memvid_sdk import use, create

# Create a new memory file
mv = use("basic", "notes.mv2", mode="auto")

# Store a document
mv.put(
    title="Project kickoff",
    label="meeting",
    metadata={"date": "2024-01-15"},
    text="Discussed timeline, assigned tasks to team members.",
)

# Search by keyword
results = mv.find("timeline")
print(results["hits"])

# Ask a question (retrieves relevant context)
answer = mv.ask("What was discussed in the kickoff?")
print(answer["context"])

# Commit changes
mv.seal()

Embeddings & semantic search

Semantic search (mode="sem") and hybrid search (mode="auto") require a vector index:

from memvid_sdk import create

mv = create("notes.mv2", enable_vec=True)

Generate embeddings during ingestion (local or OpenAI)

mv.put(
    title="Doc",
    label="note",
    metadata={},
    text="alpha fastembed test",
    enable_embedding=True,
    embedding_model="bge-small",  # local (fastembed)
)

Batch mode:

mv.put_many(
    [{"title": "Doc", "label": "note", "text": "alpha"}],
    opts={"enable_embedding": True, "embedding_model": "bge-small"},
)

Supported embedding_model values: bge-small, bge-base, nomic, gte-large, openai-small, openai-large, openai-ada.

Bring your own embeddings (precomputed)

Store embedding identity metadata so semantic queries can auto-detect the right model later:

mv.put_many(
    [{"title": "Doc", "label": "note", "text": "alpha"}],
    embeddings=[[0.1, 0.2, 0.3, 0.4]],
    embedding_identity={"provider": "custom", "model": "my-embedder-v1"},
)

Use an embedder (SDK-managed query embeddings)

from memvid_sdk.embeddings import HashEmbeddings

embedder = HashEmbeddings(dimension=32)  # deterministic offline embedder for tests
mv.put_many(
    [{"title": "Doc", "label": "note", "text": "alpha"}],
    embedder=embedder,
)

mv.find("alpha", mode="sem", embedder=embedder)
mv.ask("alpha", mode="sem", context_only=True, embedder=embedder)

Built-in embedders: OpenAIEmbeddings (OPENAI_API_KEY), CohereEmbeddings (COHERE_API_KEY), VoyageEmbeddings (VOYAGE_API_KEY), NvidiaEmbeddings (NVIDIA_API_KEY), HuggingFaceEmbeddings (local), HashEmbeddings (offline deterministic).

from memvid_sdk.embeddings import NvidiaEmbeddings

embedder = NvidiaEmbeddings(model="nvidia/nv-embed-v1")  # uses NVIDIA_API_KEY
mv.put_many([{"title": "Doc", "label": "note", "text": "alpha"}], embedder=embedder)
mv.find("alpha", mode="sem", embedder=embedder)

Auto query embeddings (no query_embedding required)

If the memory contains memvid.embedding.* identity metadata (recommended) or a known vector dimension (fallback), semantic queries can embed the query automatically:

mv.find("alpha", mode="sem")
mv.find("alpha", mode="sem", query_embedding_model="bge-small")  # override/force

Core API

Opening a memory

mv = use(kind, path, apikey=None, **options)

Parameter	Type	Description
kind	str	Adapter type: "basic", "langchain", "llamaindex", "openai", "crewai", "vercel-ai", "autogen"
path	str	Path to .mv2 file
mode	str	"open" (default), "create", or "auto"
enable_lex	bool	Enable lexical index (default: True)
enable_vec	bool	Enable vector index (default: False)
read_only	bool	Open in read-only mode (default: False)
lock_timeout_ms	int	Lock acquisition timeout in milliseconds (default: 250)
force	str | None	Set to "stale_only" to force-release stale locks
force_writable	bool	Open read-only first, then re-open writable (best-effort)

Context manager support:

with use("basic", "notes.mv2") as mv:
    mv.put(title="Note", label="general", metadata={}, text="Content")
# File handle automatically closed

Storing documents

frame_id = mv.put(
    title="Document title",
    label="category",
    metadata={"key": "value"},
    text="Document content...",
    uri="mv2://custom/path",
    tags=["tag1", "tag2"],
)

Batch ingestion

For bulk imports, put_many processes documents in parallel with 100x+ speedup over individual calls:

docs = [
    {"title": "Doc 1", "label": "news", "text": "First document..."},
    {"title": "Doc 2", "label": "news", "text": "Second document..."},
    # ... thousands more
]

frame_ids = mv.put_many(docs, opts={"compression_level": 3})
print(f"Ingested {len(frame_ids)} documents")

Searching

results = mv.find(query, k=5, snippet_chars=240, scope=None, mode=None)

Parameter	Type	Description
query	str	Search query
k	int	Number of results (default: 5)
snippet_chars	int	Snippet length (default: 240)
scope	str	Filter by URI prefix
mode	str	"auto", "lex", or "sem"

Semantic/hybrid options (when mode != "lex"):

Parameter	Type	Description
query_embedding	list[float] | None	Precomputed query embedding
query_embedding_model	str | None	Force embedding model for auto query embeddings
adaptive	bool | None	Enable adaptive retrieval cutoff
min_relevancy	float | None	Minimum relevancy (default: 0.5 when adaptive)
max_k	int | None	Maximum results (default: 100 when adaptive)
adaptive_strategy	str | None	One of: relative, absolute, cliff, elbow, combined

Retrieval-augmented generation

response = mv.ask(
    question,
    k=6,
    mode="auto",
    model="openai:gpt-4o-mini",
    api_key=os.environ.get("OPENAI_API_KEY"),
    context_only=False,
    mask_pii=False,
)

Parameter	Type	Description
question	str	Question to answer
k	int	Documents to retrieve (default: 6)
mode	str	"auto", "lex", or "sem"
model	str	LLM for synthesis (e.g., "openai:gpt-4o-mini", "nvidia:meta/llama3-8b-instruct")
api_key	str	API key for the LLM provider
context_only	bool	Skip synthesis, return context only
mask_pii	bool	Redact PII from response

Semantic/hybrid options (when mode != "lex"):

Parameter	Type	Description
query_embedding	list[float] | None	Precomputed query embedding
query_embedding_model	str | None	Force embedding model for auto query embeddings
adaptive	bool | None	Enable adaptive retrieval cutoff
min_relevancy	float | None	Minimum relevancy (default: 0.5 when adaptive)
max_k	int | None	Maximum results (default: 100 when adaptive)
adaptive_strategy	str | None	One of: relative, absolute, cliff, elbow, combined

Best practices: Adaptive retrieval

For best search quality, enable adaptive retrieval with the combined strategy. This dynamically adjusts result counts based on relevance scores rather than returning a fixed k:

# Recommended for find()
results = mv.find(
    "your query",
    mode="sem",  # or "auto"
    adaptive=True,
    adaptive_strategy="combined",
)

# Recommended for ask()
answer = mv.ask(
    "your question",
    mode="auto",
    adaptive=True,
    adaptive_strategy="combined",
)

The combined strategy uses both relative thresholds and score cliff detection to filter out low-relevance results, providing higher quality context for RAG applications.

Timeline queries

entries = mv.timeline(limit=100, since=1704067200, reverse=True)

Statistics

stats = mv.stats()
# {'frame_count': 42, 'size_bytes': 1048576, 'has_lex_index': True, ...}

File operations

# Commit pending changes
mv.seal()

# Explicit commit without sealing
mv.commit()

# Verify file integrity
report = mv.verify(deep=True)
report = Memvid.verify("notes.mv2", deep=True)  # also supported

# Repair indexes
result = mv.doctor(dry_run=True)
result = Memvid.doctor("notes.mv2", rebuild_time_index=True)

Diagnostics

from memvid_sdk import info, lock_who, lock_nudge, verify_single_file

print(info())                       # versions + enabled features
print(lock_who("notes.mv2"))        # lock status + owner (if locked)
print(lock_nudge("notes.mv2"))      # request stale lock release
verify_single_file("notes.mv2")     # ensure no sidecar files exist

Framework adapters

Adapters expose framework-native tools when the corresponding dependency is installed.

LangChain

mv = use("langchain", "notes.mv2")
tools = mv.tools  # List of StructuredTool instances

LlamaIndex

mv = use("llamaindex", "notes.mv2")
engine = mv.as_query_engine()
response = engine.query("What is the timeline?")

OpenAI function calling

mv = use("openai", "notes.mv2")
functions = mv.functions  # JSON schemas for tool_calls

CrewAI

mv = use("crewai", "notes.mv2")
tools = mv.tools  # CrewAI-compatible tools

Table extraction

Extract structured tables from PDFs:

result = mv.put_pdf_tables("report.pdf", embed_rows=True)
print(f"Extracted {result['tables_count']} tables")

tables = mv.list_tables()
data = mv.get_table(tables[0]["table_id"], format="dict")

Error handling

Typed exceptions with stable codes:

Code	Exception	Description
MV001	CapacityExceededError	Storage capacity exceeded
MV002	TicketInvalidError	Invalid ticket signature
MV003	TicketReplayError	Ticket replay detected
MV004	LexIndexDisabledError	Lexical index not enabled
MV005	TimeIndexMissingError	Time index missing
MV006	VerifyFailedError	Verification failed
MV007	LockedError	File locked by another process
MV008	ApiKeyRequiredError	API key required
MV009	MemoryAlreadyBoundError	File bound to another memory
MV010	FrameNotFoundError	Frame not found
MV011	VecIndexDisabledError	Vector index not enabled
MV012	CorruptFileError	Corrupt file / invalid TOC
MV013	FileNotFoundError	File not found
MV014	VecDimensionMismatchError	Vector dimension mismatch
MV015	EmbeddingFailedError	Embedding failed
MV016	ClipIndexDisabledError	CLIP index not enabled
MV017	NerModelNotAvailableError	NER model not available

from memvid_sdk import CapacityExceededError

try:
    mv.put(...)
except CapacityExceededError:
    print("Out of storage capacity")

Environment variables

Variable	Description
MEMVID_API_KEY	API key for capacity beyond free tier
MEMVID_API_URL	Control plane URL (for enterprise deployments)
MEMVID_OFFLINE	Set to 1 to disable network providers and model downloads
MEMVID_MODELS_DIR	Path to local embedding model cache (fastembed)
MEMVID_CLIP_MODEL	Local CLIP model name (default: mobileclip-s2)
OPENAI_API_KEY	API key for OpenAI embeddings (and can be used for ask() examples)
OPENAI_BASE_URL	Optional OpenAI-compatible base URL for embeddings
NVIDIA_API_KEY	API key for NVIDIA Integrate embeddings
NVIDIA_BASE_URL	Optional NVIDIA Integrate base URL (default: https://integrate.api.nvidia.com)
NVIDIA_EMBEDDING_MODEL	Optional NVIDIA embedding model override

Development

Build from source:

cd proprietary/memvid-bindings/python

# Create virtual environment
python -m venv .venv
source .venv/bin/activate

# Install dev dependencies
pip install maturin

# Build native extension
maturin develop --release --skip-install

# Run tests
./.venv/bin/python -m pytest -q

# Smoke-run offline-safe examples
./.venv/bin/python scripts/test_examples.py

# Optional: run gated embedding runtime tests
MEMVID_TEST_FASTEMBED=1 ./.venv/bin/python -m pytest -q tests/test_runtime_embeddings_gated.py
MEMVID_TEST_OPENAI=1 OPENAI_API_KEY=... ./.venv/bin/python -m pytest -q tests/test_runtime_embeddings_gated.py
MEMVID_TEST_NVIDIA=1 NVIDIA_API_KEY=... ./.venv/bin/python -m pytest -q tests/test_runtime_embeddings_gated.py

# Optional: run gated local model tests (requires `memvid models install ...`)
MEMVID_TEST_CLIP=1 ./.venv/bin/python -m pytest -q tests/test_phase2_models_tables.py
MEMVID_TEST_NER=1 ./.venv/bin/python -m pytest -q tests/test_phase2_models_tables.py

Requirements

• Python 3.8+
• macOS (Apple Silicon or Intel), Linux (x64), or Windows (x64)

License

Apache-2.0. See LICENSE.