Struct GraphConfig 

pub struct GraphConfig {

    pub enabled: bool,
    pub extract_model: String,
    pub max_entities_per_message: usize,
    pub max_edges_per_message: usize,
    pub community_refresh_interval: usize,
    pub entity_similarity_threshold: f32,
    pub extraction_timeout_secs: u64,
    pub use_embedding_resolution: bool,
    pub entity_ambiguous_threshold: f32,
    pub max_hops: u32,
    pub recall_limit: usize,
    pub expired_edge_retention_days: u32,
    pub max_entities: usize,
    pub community_summary_max_prompt_bytes: usize,
    pub community_summary_concurrency: usize,
    pub lpa_edge_chunk_size: usize,
    pub temporal_decay_rate: f64,
    pub edge_history_limit: usize,
    pub note_linking: NoteLinkingConfig,
    pub spreading_activation: SpreadingActivationConfig,
    pub link_weight_decay_lambda: f64,
    pub link_weight_decay_interval_secs: u64,
    pub belief_revision: BeliefRevisionConfig,
    pub rpe: RpeConfig,
    pub pool_size: u32,
}

Configuration for the knowledge graph memory subsystem ([memory.graph] TOML section).

Security

Entity names, relation labels, and fact strings extracted by the LLM are stored verbatim without PII redaction. This is a known pre-1.0 MVP limitation. Do not enable graph memory when processing conversations that may contain personal, medical, or sensitive data until a redaction pass is implemented on the write path.

Fields

enabled: bool

extract_model: String

max_entities_per_message: usize

max_edges_per_message: usize

community_refresh_interval: usize

entity_similarity_threshold: f32

extraction_timeout_secs: u64

use_embedding_resolution: bool

entity_ambiguous_threshold: f32

max_hops: u32

recall_limit: usize

expired_edge_retention_days: u32

Days to retain expired (superseded) edges before deletion. Default: 90.

max_entities: usize

Maximum entities to retain in the graph. 0 = unlimited.

community_summary_max_prompt_bytes: usize

Maximum prompt size in bytes for community summary generation. Default: 8192.

community_summary_concurrency: usize

Maximum concurrent LLM calls during community summarization. Default: 4.

lpa_edge_chunk_size: usize

Number of edges fetched per chunk during community detection. Default: 10000. Set to 0 to disable chunking and load all edges at once (legacy behavior).

temporal_decay_rate: f64

Temporal recency decay rate for graph recall scoring (units: 1/day).

When > 0, recent edges receive a small additive score boost over older edges. The boost formula is 1 / (1 + age_days * rate), blended additively with the base composite score. Default 0.0 preserves existing scoring behavior exactly.

edge_history_limit: usize

Maximum number of historical edge versions returned by edge_history(). Default: 100.

Caps the result set returned for a given source entity + predicate pair. Prevents unbounded memory usage for high-churn predicates when this method is exposed via TUI or API endpoints.

note_linking: NoteLinkingConfig

A-MEM dynamic note linking configuration.

When note_linking.enabled = true, entities extracted from each message are linked to semantically similar entities via similar_to edges. Requires an embedding store (qdrant or sqlite vector backend) to be configured.

spreading_activation: SpreadingActivationConfig

SYNAPSE spreading activation retrieval configuration.

When spreading_activation.enabled = true, graph recall uses spreading activation with lateral inhibition and temporal decay instead of BFS.

link_weight_decay_lambda: f64

A-MEM link weight decay: multiplicative factor applied to retrieval_count for un-retrieved edges each decay pass. Range: (0.0, 1.0]. Default: 0.95.

link_weight_decay_interval_secs: u64

Seconds between link weight decay passes. Default: 86400 (24 hours).

belief_revision: BeliefRevisionConfig

Kumiho AGM-inspired belief revision configuration.

When belief_revision.enabled = true, new edges that semantically contradict existing edges for the same entity pair trigger revision: the old edge is invalidated with a superseded_by pointer and the new edge becomes the current belief.

rpe: RpeConfig

D-MEM RPE-based tiered graph extraction routing.

When rpe.enabled = true, low-surprise turns skip the expensive MAGMA LLM extraction pipeline. A consecutive-skip safety valve ensures no turn is silently skipped indefinitely.

pool_size: u32

SQLite connection pool size dedicated to graph operations.

Graph tables share the same database file as messages/embeddings but use a separate pool to prevent pool starvation when community detection or spreading activation runs concurrently with regular memory operations. Default: 3.

Trait Implementations

impl Clone for GraphConfig

fn clone(&self) -> GraphConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for GraphConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for GraphConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for GraphConfig

where GraphConfig: Default,

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for GraphConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.