Scoring Model

Scoring Model

MemoryVault computes dynamic confidence and importance scores for every memory using the memoryvault-dynamic-v1 scoring model. These scores combine the user-set base values with signals derived from content analysis, graph structure, source trust, and age.

Use memory_explain_score to see the full breakdown for any memory.

How It Works

1. Start with user-provided base_confidence (default 0.7) and base_importance (default 0.5)
2. Compute deltas from multiple signal components
3. Sum base + all deltas
4. Clamp result to [0, 1]

Confidence Components

Component	Max Effect	What It Measures
source_bonus	+0.09	High-signal source keywords (api, system, official, doc, test, etc.)
source_trust_delta	+/-0.40	Trust weight set via memory_source_trust_set (0.5 = neutral)
certainty_signal	+0.20	Certainty keywords in content (verified, confirmed, exact, measured, etc.)
hedge_penalty	-0.20	Hedge keywords in content (maybe, might, perhaps, guess, probably, etc.)
type_confidence_bias	+/-0.08	Memory type: fact +0.08, journal -0.06, note 0
support_signal	+0.22	Count of supports relationship links
link_signal	+0.063	Total link count (log-scaled)
example_signal	+0.02	Count of example_of relationship links
contradiction_penalty	-0.28	Count of contradicts relationship links
source_penalty	-0.07	Low-signal source keywords (rumor, guess, hearsay, vibe)
stale_penalty	-0.20	Age > 365 days since last update

Importance Components

Component	Max Effect	What It Measures
importance_keyword_signal	+0.24	Action keywords (goal, strategy, deadline, todo, critical, priority, etc.)
source_trust_delta	+/-0.14	Source trust weight
content_depth_signal	+0.08	Content length > 80 characters (longer = more important)
type_importance_bias	+/-0.04	Memory type: note +0.04, fact +0.02, journal +0.01
link_signal	+0.18	Total link count (log-scaled)
cause_signal	+0.14	Count of causes relationship links
example_signal	+0.08	Count of example_of links
supersede_signal	+0.08	Count of supersedes links
recency_signal	+/-0.12	Age: <3 days +0.12, <14 days +0.07, <60 days +0.03, >60 days decays
contradiction_penalty	-0.07	Contradictions (scaled down from confidence impact)

Signal Detection

Keyword matching scans the concatenation of title + key + content + tags (case-insensitive).

Source trust is set per-brain via memory_source_trust_set. The baseline is 0.5 (neutral). Higher trust boosts both confidence and importance. Lower trust reduces them.

Link stats are computed from the memory_links table, counting each relationship type separately.

Age is measured from updated_at (not created_at), so updating a memory resets the age signal.

Example

A memory with:

• Type: fact, base confidence 0.7, base importance 0.5
• Content containing "verified" and "confirmed"
• 3 supports links, 1 contradicts link
• Source trust: 0.8
• Age: 5 days

Would produce roughly:

• Confidence: 0.7 + 0.08 (type) + 0.12 (source trust) + 0.08 (certainty x2) + 0.15 (supports x3) - 0.09 (contradicts) = ~1.0 (clamped)
• Importance: 0.5 + 0.02 (type) + 0.042 (source trust) + 0.18 (links) + 0.07 (recency) = ~0.81

Using Scores

• Sorting: Higher scores = more trustworthy/important memories
• Filtering: Use min_score in memory_search to filter low-confidence results
• Decay: memory_decay reduces scores for stale, poorly-connected memories
• Reinforcement: memory_reinforce boosts scores and spreads to neighbors
• Conflicts: memory_conflicts uses confidence to identify contradictions worth resolving