Struct MemoryStore 

pub struct MemoryStore {
    pub memories: Vec<MemoryRecord>,
    pub dims: usize,
    /* private fields */
}

Fields

memories: Vec<MemoryRecord>

dims: usize

Implementations

impl MemoryStore

pub fn new(dims: usize) -> Self

pub fn ensure_token_hv(&mut self, token: &str) -> &Hypervector

M1: Return reference instead of clone — eliminates 10KB clone per call.

pub fn retain(&mut self, content: &str, source: &str, error: bool) -> i64

Ingest a memory: tokenize, encode, store, update co-occurrence

pub fn encode_tokens(&mut self, tokens: &[String]) -> Hypervector

V58 P2a: deterministic token-set hypervector — majority vote over the k token HVs only (NOT chained bundle_into, which starts from an empty HV whose -1 votes poison every tie and collapse everything to zero). Sort+dedup for order-independence; FNV content seeds give distinct per-token vectors.

pub fn encode_query(&self, query: &str) -> Hypervector

Encode a query into a hypervector.

V80: this used bundle_into from an all--1 HV. bundle_into keeps the accumulator’s bit on a tie, and every tie starts as -1, so the result collapsed toward all--1 (or, once two tokens tie, stayed put) — every query scored a near-uniform 1.0 against every memory. Encode by the same deterministic majority vote as encode_tokens: for each dim set +1 iff more than half the recognised token HVs vote +1. Unknown tokens are ignored (the store has no vector for them).

pub fn recall(&self, query: &str, top_n: usize) -> Vec<ScoredMemory>

Search memories by query similarity. M4: Uses select_nth_unstable for O(N) top-K instead of O(N log N) full sort.

pub fn predict(&self, query: &str, top_n: usize) -> Vec<(String, u64)>

Co-occurrence prediction: given query tokens, predict most associated tokens

pub fn consolidate(&mut self)

Promote memories to higher tiers