Abstention threshold — collection-level gate on max text overlap.
Dense embeddings (bge-small-en-v1.5) produce 0.80+ cosine similarity even
for unrelated content, so vec_sim is NOT used for abstention.
Lowered from 0.30 to 0.15 to avoid dropping valid results for
numeric/synonym-heavy queries where word overlap is inherently lower.
Multiplicative boost for memories found via entity tag expansion.
Applied on top of the standard scoring pipeline. AutoMem uses +0.15 additive;
we use a multiplicative 1.15 to integrate with the existing RRF-based scoring.
Graph enrichment — re-enabled at a conservative factor.
Grid search over [0.0, 0.05, 0.1, 0.15, 0.2, 0.3] showed no regression on
LoCoMo (benchmark data has sparse relationship graphs). 0.1 chosen as a
safe default: neighbors get at most 10 % of the seed score, enough to
surface related memories in production without dominating rankings.
Feedback is an explicit user/system signal — asymmetric by design.
Negative feedback aggressively suppresses (explicit downvote).
Positive feedback gives only a mild boost (prevents displacing unrelated results).