Skip to content

Types

Public dataclasses, Protocols, and type aliases. Everything here is re-exported by mneme.__init__.

Dataclasses

Hit dataclass 

Hit(response: str, similarity: float, confidence: float, age_seconds: int, layer: HitLayer, namespace: str, metadata: dict[str, Any])

Stats dataclass 

Stats(namespace: str | None, entries: int, hits_exact: int, hits_semantic: int, misses: int, evictions: int, expirations: int, embedder_fingerprint: str, vector_dtype: str, memory_bytes_estimate: int, index_memory_bytes: int | None = None, index_tombstone_count: int | None = None)

Health dataclass 

Health(healthy: bool, schema_version: int, integrity_ok: bool, embedder_fingerprint_match: bool, entries: int, namespaces: int, oldest_entry_age_seconds: int | None, index_backend: str, store_backend: str, vector_dtype: str, multi_process_mode: str)

StoredEntry dataclass 

StoredEntry(id: int, namespace: str, query_hash: str, query: str, response: str, embedding: bytes, metadata: dict[str, Any], created_at: int, last_accessed_at: int, ttl: int | None, access_count: int)

Protocols

Embedder

Bases: Protocol

Sync embedder. Returns 1-D float32 vectors of length dim.

AsyncEmbedder

Bases: Protocol

Async embedder. Returns 1-D float32 vectors of length dim.

Index

Bases: Protocol

In-memory vector index over L2-normalized float32 vectors.

Store

Bases: Protocol

Persistent backend. Source of truth for entries; vectors persisted as float32.

Implementations must provide write atomicity for insert (vector and metadata committed together) and durable reads for get_by_hash.

open 

open(embedder_fingerprint: str, embedder_dim: int) -> None

Initialize storage. Validate fingerprint/dim against any existing data; raise EmbedderMismatchError or EmbedderDimensionError on mismatch.

insert 

insert(entry: StoredEntry) -> int

Returns the assigned id and bumps version_counter in the same txn.

snapshot_to 

snapshot_to(dest_path: str | Path) -> None

Copy the entire store to dest_path. Format is implementation-defined.

Sync ↔ async embedder adapters

to_async_embedder 

to_async_embedder(embedder: Embedder) -> AsyncEmbedder

Wrap a sync Embedder so it satisfies AsyncEmbedder.

to_sync_embedder 

to_sync_embedder(embedder: AsyncEmbedder) -> Embedder

Wrap an AsyncEmbedder so it satisfies Embedder (sync).

Type aliases

mneme exports four Literal aliases for configuration keys:

  • VectorDtype - "float32" | "float16" | "int8". The in-memory matrix dtype.
  • HitLayer - "exact" | "semantic". Which layer answered a get.
  • IndexBackend - "numpy" | "hnsw" | "auto". The index implementation.
  • MultiProcessMode - "single" | "stale-tolerant" | "mmap-shared". Coordination across processes.

Plus three callable aliases:

  • ConfidenceFn - Callable[[float, int, dict[str, Any]], float] - your confidence scorer.
  • Validator - Callable[[str], bool] - your response validator.
  • MetricsHook - Callable[[str, dict[str, Any]], None] - your metrics fan-out.