All settings live on AriadneConfig. Every field has a default — you can construct AriadneMemory() with none of them — but db_path and embedding_dim are the two you'll usually set.

AriadneConfig

from arriadne import AriadneConfig, AriadneMemory

config = AriadneConfig(
    db_path="memory.db",          # SQLite database path
    embedding_dim=384,            # Vector dimensions (must match your embedder)
    faiss_type="auto",            # "auto" | "flat_ip" | "ivf_flat"
    ivf_threshold=50_000,         # auto: switch Flat -> IVF at this many vectors
    ivf_nlist=128,                # IVF Voronoi cells (capped at sqrt(n))
    ivf_min_points=1_000,         # ivf_flat: min vectors before IVF is trained
    dedup_threshold=0.8,          # MinHash Jaccard threshold for near-dupes
    dedup_num_perm=128,           # MinHash permutations
    consolidation_threshold=0.7,  # Jaccard threshold for consolidation grouping
    consolidation_min_group=2,    # Min memories to form a consolidation group
    eviction_budget=0.1,          # Fraction of memories evicted per evict() run
    retention_half_life=86_400.0, # Ebbinghaus half-life in seconds (1 day)
    retention_growth_factor=1.5,  # Stability multiplier applied on each access
    retention_strength_cap=100.0, # Ceiling for accrued retention strength
    priority_weights={
        "importance": 0.4,
        "recency": 0.3,
        "access_count": 0.2,
        "retention": 0.1,
    },
    max_graph_depth=10,           # Hard cap on graph traversal hops
    max_access_log_per_memory=50, # Access-log rows kept per memory after pruning
    wal_autocheckpoint=1000,      # SQLite WAL autocheckpoint (pages)
)

mem = AriadneMemory(config=config)

Defaults

Setting	Default	Description
db_path	"arriadne.db"	SQLite database file path
embedding_dim	384	Vector dimension (matches all-MiniLM-L6-v2)
faiss_type	"auto"	"auto", "flat_ip", or "ivf_flat"
ivf_threshold	50000	auto mode: vectors before upgrading Flat → IVF
ivf_nlist	128	IVF cells; effective nlist = min(ivf_nlist, √n)
ivf_min_points	1000	ivf_flat mode: vectors before the IVF index trains
dedup_threshold	0.8	MinHash Jaccard threshold for near-duplicates
dedup_num_perm	128	MinHash permutations
consolidation_threshold	0.7	Jaccard threshold for consolidation grouping
consolidation_min_group	2	Minimum group size to consolidate
eviction_budget	0.1	Fraction of active memories evicted per run
retention_half_life	86400.0	Ebbinghaus half-life, seconds
retention_growth_factor	1.5	Stability multiplier per access (≥ 1.0)
retention_strength_cap	100.0	Max accrued retention strength
max_graph_depth	10	Maximum graph traversal hops
max_access_log_per_memory	50	Access-log rows kept per memory after pruning
wal_autocheckpoint	1000	SQLite WAL autocheckpoint interval (pages)

AriadneConfig validates its inputs and raises ValueError for out-of-range values (e.g. embedding_dim < 1, a dedup_threshold outside [0, 1], an eviction_budget outside (0, 1], or an unknown faiss_type).

Embedder

The embedding model is passed to AriadneMemory, not stored on the config — when set, remember() and recall() embed text automatically:

from arriadne import AriadneMemory
from arriadne.embeddings import SentenceTransformerEmbedder

embedder = SentenceTransformerEmbedder("all-MiniLM-L6-v2")
mem = AriadneMemory(db_path="memory.db", embedding_dim=embedder.dim, embedder=embedder)

You can also pass any callable str -> list[float] (with a .dim attribute) or a model-name string. Without an embedder, Ariadne is a keyword store unless you pass vectors to remember/recall yourself. See Embeddings.