Core Concepts

Importance Scores

MemorySync separates "how important is this to the user" from "how reliable is this row" from "how recent is this content". Each gets its own column. This page is the score-by-score breakdown.

Why there are two importance columns

Field	Type / range	Authoritative source
`importance`	float, default 0.5, range 0..1	User payload or semantic engine.
`importance_score`	int, nullable, 0..100	Deterministic post-commit scorer (no LLM).

The float is for soft, user-side weighting. The integer is for the platform's own signal — it is computed cost-aware and skipped silently if the scorer fails.

Quality, confidence, and freshness

Field	Range	Computed by
`quality_score`	0..1	Semantic engine — does the text carry real information?
`confidence_score`	0..1	Combination of evidence, consistency, and freshness.
`freshness_score`	0..1	Content-type-aware decay.

Evidence and consistency

These two are graph-level signals — they describe how a row relates to its neighbours.

Field	Default	How it changes
`evidence_count`	0	Increases as corroborating relationships are recorded.
`consistency_score`	1.0	Decreases per contradiction edge against the row.

Decay and usage

Two more signals describe time and traffic.

Field	Default	How it changes
`decay_score`	1.0	Decreases over time when the row is not recalled.
`usage_count`	0	Increases each time the row contributes to a query result.

How scores combine in retrieval

The recall engine blends vector similarity with these signals. Higher importance, importance_score, quality_score, confidence_score, freshness_score, and usage_count push a row up; high decay_score keeps a row in play; low consistency_score pushes a contradicted row down.

Reading scores back from the API

BASH

curl -X POST https://api.memorysync.io/memory/query \
  -H "X-API-Key: $MEMORYSYNC_API_KEY" \
  -H "X-End-User-ID: u_42" \
  -H "X-Project-ID: $MEMORYSYNC_PROJECT_ID" \
  -H "Content-Type: application/json" \
  -d '{ "query": "what does this user prefer?", "limit": 3 }'

The response includes importance, importance_score, quality_score, confidence_score, freshness_score, evidence_count, consistency_score, usage_count, and decay_score. Use them as ranking signals or as filters for downstream agents.

What not to do with scores

Do not write importance_score, quality_score, confidence_score, or freshness_score from the client — the server overwrites them.
Do not interpret importance as a probability — it is a soft weight.
Do not use decay_score as a cache invalidation signal — it is not monotonic in real time.

← Previous

Memory Tiers

Source Provenance