Scoring

Hybrid retrieval ranks memories using a weighted score:

final = semantic * w_semantic + recency * w_recency + importance * w_importance

Default weights (if unspecified):

• semantic: 0.7
• recency: 0.2
• importance: 0.1

Importance Score

importanceScore is a 0..1 value on each memory. Higher values boost ranking and influence lifecycle behavior (for example distillation decisions in episodic memory).

Recency Score

Recency decays over time. A common interpretation is exponential decay with a 7-day half-life:

recency(t) = 0.5 ^ (age_days / 7)

This preserves recent relevance without discarding older high-signal memory.

Semantic Score

Semantic relevance is computed from embedding similarity (cosine-style similarity from nearest-neighbor vector search).

Hybrid Ranking in Practice

When you query POST /memories/query or db.memories.search(...):

• semantic match finds related content
• recency favors newer events
• importance keeps high-value memory ranked

Tune weights per use case with weights in query options.