The graph learns from use, Hebbian style: edges that keep firing together cross…

Max Kle1nz@maxkle1nz

The graph learns from use, Hebbian style: edges that keep firing together cross the LTP threshold and get a +0.15 bonus; paths that keep disappointing cross LTD and lose the same. The interesting constants are the guardrails — HOMEOSTATIC_CEILING (5.0) stops a hot path from monopolizing every future query, WEIGHT_FLOOR (0.05) keeps an edge from decaying into deletion, and CAS_RETRY_LIMIT makes weight updates safe under concurrency. Open question I keep revisiting: whether LTP/LTD thresholds should adapt per-repo instead of being global.

pub const DEFAULT_LEARNING_RATE: f32 = 0.08;
pub const DEFAULT_DECAY_RATE: f32 = 0.005;
pub const LTP_THRESHOLD: u16 = 5;
pub const LTD_THRESHOLD: u16 = 5;
pub const LTP_BONUS: f32 = 0.15;
pub const LTD_PENALTY: f32 = 0.15;
pub const HOMEOSTATIC_CEILING: f32 = 5.0;
pub const WEIGHT_FLOOR: f32 = 0.05;
pub const WEIGHT_CAP: f32 = 3.0;
/// Default ring buffer capacity for query memory (FM-PL-005).
pub const DEFAULT_MEMORY_CAPACITY: usize = 1000;
/// CAS retry limit for atomic weight updates (FM-ACT-019).
pub const CAS_RETRY_LIMIT: u32 = 64;