Skip to content

0001 — Power-law decay rejected

Status: Accepted Date: 2026-05-23 Deciders: elf (curator), Ben

Context

docs/plans/plan_memory_scoring.md proposed shipping power-law retrievability decay as an opt-in v0.17 feature, behind a memory.use_power_law_decay: bool = False flag. The proposal cited FSRS-5 research showing power-law beats exponential for flashcard-style spaced repetition.

The research agent (cited in plan_memory_scoring.md) graded this evidence A for flashcards but D for agent memory — the empirical case had not been made in our domain.

Alternatives considered

  1. Ship as default: replace exp(-λt) with (1 + 0.5t/stability)^(-0.5) for all users
  2. Ship as opt-in flag (the original plan): expose the formula behind a config flag for users to A/B
  3. Reject entirely: don't ship the mechanism at all
  4. Defer: wait for more research

Decision

Reject entirely. Drop from v0.17 plans. Document the empirical refutation here so the decision is searchable.

Evidence

Simulation (scripts/longitudinal_sim/mc_scoring_proposed.py) tested four scenarios across 4 strategies × 2 seeds. Power-law decay (v017_full) underperformed across every scenario:

scenario          v016 (exponential)   v017_full (power-law)
─────────────────────────────────────────────────────────────
baseline             80.7%                 75.3%   (−5.4pp)
weekly_rescore       81.4%                 76.4%   (−5.0pp)
long_horizon         72.4%                 72.3%   (essentially flat)
uncertain_mix        78.4%                 70.7%   (−7.6pp)

Recent reach (plasticity) catastrophically degraded:

scenario          v016                  v017_full
─────────────────────────────────────────────────
baseline             78.4%                 25.2%   (−53pp)
weekly_rescore       78.0%                 13.6%   (−64pp)
long_horizon         64.0%                 28.4%   (−36pp)
uncertain_mix        76.8%                 33.6%   (−43pp)

Mechanism: at t=1000 hours, exponential gives recency = 4.5×10⁻⁵; power-law gives 0.41. A 10,000-fold difference at the long-time end. Power-law makes one-year-old blocks competitive against one-day-old blocks. Top-K fills with stale content.

Consequences

  • v0.17 ships exp(-λt) decay only. No flag, no opt-in.
  • Documentation in plan_memory_scoring.md should add this ADR's reference under "What we are explicitly NOT adopting."
  • Future contributors should consult this ADR before reproposing power-law.

Trigger to revisit

Only if MemoryAgentBench (or equivalent agent-memory benchmark) shows power-law wins on agent workloads with statistical significance. None reported as of 2026-05.

References