Configure OSS Stack

All MemWire behaviour is controlled through MemWireConfig. Choose your vector store, embedding model, and LLM provider, then tune recall and graph settings to fit your use case.

from memwire import MemWire, MemWireConfig

config = MemWireConfig(
    qdrant_path="./memwire_data",
    qdrant_collection_prefix="app_",
)
memory = MemWire(config=config)

Components

Vector Databases

Connect to a vector store.

Embedding Models

Choose embedding model.

LLMs

Use any LLM provider with recalled context.

SQL Databases

Choose where to store metadata ledger for memories, graph edges, and categories.

Recall tuning

Recall assembles relevant memories into a formatted context string for injection into LLM prompts.

config = MemWireConfig(
    recall_min_relevance=0.25,   # minimum similarity score to include a memory
    recall_max_paths=10,         # max memory paths returned
    recall_max_depth=4,          # BFS depth when traversing the memory graph
    recall_seed_top_k=5,         # top-k seed nodes before graph traversal
    tension_threshold=0.6,       # similarity above which two memories are flagged as contradictory
    recency_weight=0.3,          # weight of recency vs relevance in scoring
    recency_halflife=3600.0,     # half-life in seconds for recency decay
)

Parameter	Default	Description
`recall_min_relevance`	`0.25`	Minimum cosine similarity for a memory to be included in recall.
`recall_max_paths`	`10`	Maximum number of memory paths returned per recall call.
`recall_max_depth`	`4`	Maximum BFS depth when traversing the memory graph.
`recall_seed_top_k`	`5`	Number of top-k seed nodes to start graph traversal from.
`recall_bfs_max_branch`	`5`	Maximum branches per node during BFS traversal.
`recall_bfs_max_paths`	`200`	Maximum candidate paths evaluated before pruning.
`tension_threshold`	`0.6`	Similarity above which two memories are flagged as contradictory.
`recency_weight`	`0.3`	Weight given to recency vs semantic relevance when scoring memories.
`recency_halflife`	`3600.0`	Half-life in seconds for recency decay. Lower = faster decay.

Graph construction

MemWire organises memories in a graph. Edges encode semantic relationships between memories.

config = MemWireConfig(
    displacement_threshold=0.15,
    node_merge_similarity=0.85,
    edge_weight_default=0.5,
    edge_decay_rate=0.02,
    edge_reinforce_amount=0.1,
)

Parameter	Default	Description
`displacement_threshold`	`0.15`	Minimum similarity required to create an edge between two memories.
`node_merge_similarity`	`0.85`	Similarity above which two memory nodes are merged (deduplication).
`edge_weight_default`	`0.5`	Initial weight for new edges.
`edge_weight_min`	`0.01`	Floor for edge weights after decay.
`edge_weight_max`	`1.0`	Ceiling for edge weights after reinforcement.
`edge_decay_rate`	`0.02`	Rate at which unused edges decay per cycle.
`edge_reinforce_amount`	`0.1`	Amount added to an edge weight on positive feedback.
`cross_memory_recent_limit`	`50`	Number of recent memories considered when building cross-memory edges.

Memory classification

Memories are automatically classified into categories using zero-shot cosine similarity.

Category	Description
`fact`	Factual statements or pieces of information
`preference`	Personal preferences or opinions
`instruction`	Directives or rules to follow
`event`	Things that happened
`entity`	Information about a person, place, or organisation

config = MemWireConfig(
    classification_threshold=0.05,
    default_anchors={
        "fact": ["This is a factual statement"],
        "preference": ["This is a personal preference"],
        "instruction": ["This is a rule or directive to follow"],
        "event": ["This is something that happened"],
        "entity": ["This is about a specific person or organisation"],
    }
)

Feedback loop

config = MemWireConfig(
    feedback_strengthen_rate=0.1,
    feedback_weaken_rate=0.05,
    feedback_align_strengthen=0.5,
    feedback_align_weaken=0.2,
)

Parameter	Default	Description
`feedback_strengthen_rate`	`0.1`	Amount to strengthen edges that contributed to a good response.
`feedback_weaken_rate`	`0.05`	Amount to weaken edges that did not contribute.
`feedback_align_strengthen`	`0.5`	Alignment score threshold above which edges are strengthened.
`feedback_align_weaken`	`0.2`	Alignment score threshold below which edges are weakened.

SQL Databases

MemWire stores memory metadata — content, graph edges, categories, and access counts — in a SQL database, while all embedding vectors live separately in a vector store. SQLite is the default because it requires no server or configuration, letting you run MemWire locally out of the box; switch to PostgreSQL or any other SQLAlchemy-compatible database by setting database_url.

config = MemWireConfig(
    database_url="sqlite:///memwire.db",
    org_id="default",
)

Parameter	Default	Description
`database_url`	`None`	SQLAlchemy-compatible URL. Defaults to `sqlite:///memwire_{org_id}.db`.
`org_id`	`"default"`	Organisation identifier for multi-tenant isolation.

Performance

config = MemWireConfig(
    background_threads=2,
    embedding_cache_maxsize=10000,
)

Parameter	Default	Description
`background_threads`	`2`	Background threads for async graph and storage writes.
`embedding_cache_maxsize`	`10000`	LRU cache size for embedding vectors.

from memwire import MemWire, MemWireConfig

config = MemWireConfig(
    qdrant_path="./memwire_data",
    qdrant_collection_prefix="app_",
)
memory = MemWire(config=config)

No environment variables are required — every setting is an explicit Python argument with a sensible default.

Search quality

Control how memories are retrieved during search and recall.

config = MemWireConfig(
    use_hybrid_search=True,    # combine dense + sparse vectors
    use_reranking=False,       # apply a cross-encoder reranker
    reranker_model_name="Xenova/ms-marco-MiniLM-L-6-v2",
)

Parameter	Default	Description
`use_hybrid_search`	`True`	Combine dense and sparse vectors for more accurate retrieval.
`use_reranking`	`False`	Apply a cross-encoder reranker to re-score top candidates.
`reranker_model_name`	`Xenova/ms-marco-MiniLM-L-6-v2`	Reranker model (used when `use_reranking=True`).

Recall tuning

Recall assembles relevant memories into a formatted context string for injection into LLM prompts.

config = MemWireConfig(
    recall_min_relevance=0.25,   # minimum similarity score to include a memory
    recall_max_paths=10,         # max memory paths returned
    recall_max_depth=4,          # BFS depth when traversing the memory graph
    recall_seed_top_k=5,         # top-k seed nodes before graph traversal
    tension_threshold=0.6,       # similarity above which two memories are flagged as contradictory
    recency_weight=0.3,          # weight of recency vs relevance in scoring
    recency_halflife=3600.0,     # half-life in seconds for recency decay
)

Parameter	Default	Description
`recall_min_relevance`	`0.25`	Minimum cosine similarity for a memory to be included in recall.
`recall_max_paths`	`10`	Maximum number of memory paths returned per recall call.
`recall_max_depth`	`4`	Maximum BFS depth when traversing the memory graph.
`recall_seed_top_k`	`5`	Number of top-k seed nodes to start graph traversal from.
`recall_bfs_max_branch`	`5`	Maximum branches per node during BFS traversal.
`recall_bfs_max_paths`	`200`	Maximum candidate paths evaluated before pruning.
`tension_threshold`	`0.6`	Similarity above which two memories are flagged as contradictory.
`recency_weight`	`0.3`	Weight given to recency vs semantic relevance when scoring memories.
`recency_halflife`	`3600.0`	Half-life in seconds for recency decay. Lower = faster decay.

Graph construction

MemWire organises memories in a graph. Edges encode semantic relationships between memories.

config = MemWireConfig(
    displacement_threshold=0.15,  # min similarity to create an edge
    node_merge_similarity=0.85,   # similarity above which two nodes are merged (dedup)
    edge_weight_default=0.5,
    edge_weight_min=0.01,
    edge_weight_max=1.0,
    edge_decay_rate=0.02,         # edges decay over time when not reinforced
    edge_reinforce_amount=0.1,    # amount edges are boosted on positive feedback
)

Parameter	Default	Description
`displacement_threshold`	`0.15`	Minimum similarity required to create an edge between two memories.
`node_merge_similarity`	`0.85`	Similarity above which two memory nodes are merged (deduplication).
`edge_weight_default`	`0.5`	Initial weight for new edges.
`edge_weight_min`	`0.01`	Floor for edge weights after decay.
`edge_weight_max`	`1.0`	Ceiling for edge weights after reinforcement.
`edge_decay_rate`	`0.02`	Rate at which unused edges decay per cycle.
`edge_reinforce_amount`	`0.1`	Amount added to an edge weight on positive feedback.
`cross_memory_recent_limit`	`50`	Number of recent memories considered when building cross-memory edges.

Memory classification

Memories are automatically classified into categories. Classification uses zero-shot cosine similarity against anchor phrases.

Category	Description
`fact`	Factual statements or pieces of information
`preference`	Personal preferences or opinions
`instruction`	Directives or rules to follow
`event`	Things that happened
`entity`	Information about a person, place, or organisation

config = MemWireConfig(
    classification_threshold=0.05,  # minimum score margin between top categories
)

You can override the anchor phrases for each category:

config = MemWireConfig(
    default_anchors={
        "fact": ["This is a factual statement"],
        "preference": ["This is a personal preference"],
        "instruction": ["This is a rule or directive to follow"],
        "event": ["This is something that happened"],
        "entity": ["This is about a specific person or organisation"],
    }
)

Feedback loop

The feedback loop reinforces graph edges that led to good responses, and weakens edges that did not contribute.

config = MemWireConfig(
    feedback_strengthen_rate=0.1,    # edge boost on positive feedback
    feedback_weaken_rate=0.05,       # edge reduction on negative feedback
    feedback_align_strengthen=0.5,
    feedback_align_weaken=0.2,
)

Parameter	Default	Description
`feedback_strengthen_rate`	`0.1`	Amount to strengthen edges that contributed to a good response.
`feedback_weaken_rate`	`0.05`	Amount to weaken edges that did not contribute.
`feedback_align_strengthen`	`0.5`	Alignment score threshold above which edges are strengthened.
`feedback_align_weaken`	`0.2`	Alignment score threshold below which edges are weakened.

Call memory.feedback() after each LLM response to close the loop:

result = memory.recall("How should I format my answers?", user_id="alice")
# ... call your LLM with result.formatted as context ...
memory.feedback(response="<llm response here>", user_id="alice")

Storage

config = MemWireConfig(
    database_url="sqlite:///memwire.db",  # SQLAlchemy-compatible URL
    org_id="default",                     # organisation identifier for multi-tenant isolation
)

Parameter	Default	Description
`database_url`	`None`	SQLAlchemy-compatible database URL. Defaults to `sqlite:///memwire_{org_id}.db`.
`org_id`	`"default"`	Organisation identifier. Used as a namespace for multi-tenant deployments.

Performance

config = MemWireConfig(
    background_threads=2,         # threads for async background writes
    embedding_cache_maxsize=10000, # LRU cache for embedding vectors
)

Parameter	Default	Description
`background_threads`	`2`	Number of background threads for async graph and storage writes.
`embedding_cache_maxsize`	`10000`	LRU cache size for embedding vectors. Increase for large workloads.

Getting started

Features

Configuration

Configure OSS Stack

Components

Vector Databases

Embedding Models

LLMs

SQL Databases

Recall tuning

Graph construction

Memory classification

Feedback loop

SQL Databases

Performance

Search quality

Recall tuning

Graph construction

Memory classification

Feedback loop

Storage

Performance

​Components

Vector Databases

Embedding Models

LLMs

SQL Databases

​Recall tuning

​Graph construction

​Memory classification

​Feedback loop

​SQL Databases

​Performance

​Search quality

​Recall tuning

​Graph construction

​Memory classification

​Feedback loop

​Storage

​Performance

Components

Recall tuning

Graph construction

Memory classification

Feedback loop

SQL Databases

Performance

Search quality

Recall tuning

Graph construction

Memory classification

Feedback loop

Storage

Performance