Architecture · Ranking

Scoring and Ranking Memories by Relevance

Memory scoring ranks candidate memories by relevance, recency and importance so agents surface the right facts — not just the most similar embeddings — at each turn.

Ranking formula

R
Relevance
Similarity
T
Recency
Time decay
I
Importance
Salience

Problem

Why similarity alone fails

Vector similarity measures embedding distance — not whether a memory is useful, current or scoped to the right user.

  • Stale but similar — “user mentioned rain last Tuesday” ranks high for a dinner query
  • Wrong scope — another user’s memory leaks without user_id filters
  • Low-salience match — small talk outranks core preferences on paraphrase overlap

Memory retrieval

Signals

Scoring signals

Production rankers combine semantic similarity, recency decay, importance boosts, memory-type weights and graph distance.

Park et al. (2023, Generative Agents): score = recency + importance + relevance (weighted sum, each normalized 0–1). Tune weights per use case — support bots may weight recency higher; assistants may weight importance for stable prefs.

SignalWhat it measuresTypical source
RelevanceQuery–memory similarityEmbedding cosine
RecencyHow fresh the memory isExponential time decay
ImportanceSalience at write timeLLM score or heuristic
Type weightEpisodic vs semantic priorityMetadata tag
Graph distanceEntity relationship hopsZep Graphiti traversal

Recency

Recency decay

Memories lose rank as they age — exponentially, with a tunable half-life.

Park et al. (2023) use decay at 0.995 per hour since last access. Worked example: a fresh preference (recency 0.9) outranks week-old small talk (recency 0.3) when relevance is comparable. Hard-delete stale memories via forgetting and eviction; downrank when you may still need archival access.

Forgetting and eviction

Hybrid

Hybrid ranking with keyword and metadata

Pre-filter by user_id, session and tags — then combine BM25 keyword scores with vector similarity.

Hybrid search catches exact order numbers and SKUs vectors miss. Metadata filters prevent cross-user leakage before ranking runs.

Hybrid search for memory retrieval

Frameworks

Framework scoring approaches

  • Engram — Weaviate hybrid search + metadata filters on memory collections
  • Mem0 — relevance params on search API; LOCOMO J 66.9 with tuned retrieval (Chhikara et al., 2025)
  • Zep — graph traversal + vector hybrid; temporal validity filters
  • Custom — cross-encoder rerankers on top-k candidates (adds latency, improves precision)

Best AI memory tools

FAQ

Frequently asked questions

How do you choose top-k for memory scoring?

Start with k = 5–10, tune on LOCOMO or your domain eval. Smaller k saves tokens; larger k improves recall. Mem0's selective retrieval uses far fewer tokens than full-context baselines (Chhikara et al., 2025).

Recency vs relevance — which to weight higher?

Depends on use case: support bots often weight recency for ticket context; personalization assistants weight importance for stable prefs. Park et al. (2023) uses equal 1.0 weights as a starting point.

Is a cross-encoder reranker worth it?

Yes when precision matters and you can afford extra latency — rerank top-20 vector hits down to top-5. Skip for latency-sensitive paths; metadata filters + recency often suffice.

How do you fix wrong memory retrieval?

Add recency/importance to scoring, tighten user_id filters, use hybrid search for exact IDs, invalidate stale facts. See scoring signals above.

How do you benchmark memory scoring?

LOCOMO and LongMemEval measure end-to-end recall. Ablate scoring weights and track hit rate + tokens injected. See evaluation hub.

How does scoring relate to context engineering?

Scoring decides which memories enter the prompt; context engineering formats and orders them under token budget. See context engineering.

What is the Park et al. memory scoring formula?

Weighted sum of recency + importance + relevance (each 0–1), with recency decaying at 0.995 per hour since last access (Park et al., 2023, Generative Agents).

Does Engram support hybrid memory scoring?

Yes — Weaviate hybrid search combines vector similarity with keyword/BM25 and metadata filters on memory collections. See Engram explained.