Memory Worth

A memory recalled 100 times but never leading to a good outcome is noise. A memory recalled three times but consistently leading to wins is gold. Frequency ≠ utility.

Memory Worth is a per-row signal that converges to the true utility of a memory over time. It’s how Mnemosyne decides what to promote, what to demote, what to merge, and what to forget.

The two counters

Every memorable row carries:

worth_success NUMERIC DEFAULT 0,   -- recalled AND outcome was good
worth_failure NUMERIC DEFAULT 0,   -- recalled AND outcome was bad

The counters are fractional (NUMERIC), not integers — credit propagation writes partial weights. The stored worth column is the raw ratio success / (success + failure), defaulting to 0.5 before any observations.

worth = worth_success / (worth_success + worth_failure)   -- default 0.5

A Wilson score interval lower bound is computed app-side when ranking needs a confidence-adjusted score — it handles cold-start so a memory with 1 success and 0 failures isn’t promoted above one with 10 successes and 1 failure just because the raw ratio is 1.0:

wilson ≈ (p̂ + z²/2n − z√(p̂(1−p̂)/n + z²/4n²)) / (1 + z²/n)

(p̂ = success ratio, n = total observations, z = 1.96)

Decision rule

worth > 0.7  AND  confidence > 0.5   →  promote (boost activation, candidate to pin)
worth < 0.3  AND  confidence > 0.5   →  demote (lower activation, archive candidate)
confidence  < 0.3                    →  insufficient data, neutral treatment

After ~10 observations the signal converges toward ground-truth utility.

How worth updates

Three sources, weighted by signal confidence:

The heuristic always fires (zero cost). Explicit signals move worth in five observations; heuristic-only takes twenty-five. That’s deliberate — you need high-quality evidence to move scores fast.

Credit propagation (TD-λ)

When a memory’s outcome is recorded, the credit propagates backwards through the provenance DAG — facts that fed decisions that fed strategies that succeeded all get partial credit, decaying by λ at each hop:

Strategy succeeded → +1.0 direct
Decision that fed it → +0.7 (λ = 0.7)
Fact that derived the decision → +0.49 (λ²)

This uses the existing mnemo_relation edges (derived_from, part_of). You don’t need to track provenance separately — it’s already in the graph.

What worth feeds

• Hot tier eligibility — only worth ≥ 0.5 items get promoted to the Redis activation set.
• Consolidation gate — high-worth memories are exempt from merging (they’re useful as-is, don’t summarise them away). See Consolidation gate.
• Recall ranking — final score blends cosine × recency × worth.
• Self-edit safety — strengthen/weaken actions cap at a few per day to prevent gaming.