Configuration¶

All configuration lives in MemoryConfig. Pass it to Memory() directly, or override individual parameters.

MemoryConfig¶

Python

from memv import Memory, MemoryConfig

config = MemoryConfig(
    max_statements_for_prediction=5,
    enable_episode_merging=False,
)
memory = Memory(config=config, embedding_client=embedder, llm_client=llm)

Individual params override config values:

Python

# config says auto_process=False, but this overrides it
memory = Memory(config=config, auto_process=True, ...)

Reference¶

Database¶

Parameter	Default	Description
`db_path`	`".db/memory.db"`	SQLite database file path. Parent directories are auto-created.
`embedding_dimensions`	`1536`	Vector dimensions. Must match your embedding model.

Processing Triggers¶

Parameter	Default	Description
`auto_process`	`False`	Enable automatic background processing when message threshold is reached.
`batch_threshold`	`10`	Number of messages that triggers auto-processing.
`max_retries`	`1`	Retry attempts on processing failure.

Segmentation¶

Parameter	Default	Description
`segmentation_threshold`	`20`	Maximum messages per episode.
`time_gap_minutes`	`30`	Time gap (minutes) that triggers a new episode boundary.

Episode Merging¶

Parameter	Default	Description
`enable_episode_merging`	`True`	Merge similar episodes to reduce redundancy.
`merge_similarity_threshold`	`0.9`	Embedding similarity threshold for merging (0-1).

Knowledge Deduplication¶

Parameter	Default	Description
`enable_knowledge_dedup`	`True`	Deduplicate similar knowledge entries.
`knowledge_dedup_threshold`	`0.8`	Similarity threshold for deduplication (0-1).

Predict-Calibrate¶

Parameter	Default	Description
`max_statements_for_prediction`	`10`	How many existing knowledge statements to use during prediction. More = better predictions, higher token cost.

Retrieval¶

Parameter	Default	Description
`search_top_k_episodes`	`10`	Default max episodes returned per query.
`search_top_k_knowledge`	`10`	Default max knowledge entries returned per query.

Embedding Cache¶

Parameter	Default	Description
`enable_embedding_cache`	`True`	Cache query embeddings to reduce API calls.
`embedding_cache_size`	`1000`	Max entries in the LRU cache.
`embedding_cache_ttl_seconds`	`600`	Cache entry TTL (10 minutes).

Common Configurations¶

High-throughput chat agent¶

Python

memory = Memory(
    auto_process=True,
    batch_threshold=20,
    enable_episode_merging=True,
    enable_knowledge_dedup=True,
    enable_embedding_cache=True,
    # ...
)

Precision extraction (slower, more accurate)¶

Python

memory = Memory(
    max_statements_for_prediction=20,
    knowledge_dedup_threshold=0.95,
    merge_similarity_threshold=0.95,
    # ...
)

Minimal (no optional processing)¶

Python

memory = Memory(
    enable_episode_merging=False,
    enable_knowledge_dedup=False,
    enable_embedding_cache=False,
    # ...
)