Weibull Decay

Mnemo is the first AI memory framework to use a cognitive science-based forgetting model.

Why Forgetting Matters

Other memory systems store everything forever. This leads to:

• Noise accumulation — outdated facts drown relevant ones
• Contradictions — old preferences override new ones
• Slow retrieval — searching through irrelevant memories

Mnemo models human forgetting using the Weibull distribution, the same model used in reliability engineering and cognitive psychology.

How It Works

The Weibull stretched-exponential function controls how memories fade over time. Each memory tier has its own shape parameter (β), optimized through 35 ablation tests to match cognitive science research on different memory types:

• Core memories (user identity, key preferences) — sub-exponential decay: slow to fade, highly persistent
• Working memories (recent conversations, active tasks) — standard exponential decay
• Peripheral memories (one-off mentions, trivia) — super-exponential: fades faster than standard

Important memories with high importance or emotionalSalience decay more slowly. Frequently accessed memories are reinforced — just like human recall strengthens neural pathways.

Memory Tiers

Different memories decay at different rates, just like in human cognition:

Tier	Behavior	Example
Core	Very persistent, almost never forgotten	User's name, job, key preferences
Working	Standard decay rate	Recent conversations, tasks
Peripheral	Fades quickly unless reinforced	One-off mentions, trivia

Tier Transitions

Memories move between tiers based on usage patterns. Promotion and demotion are driven by a composite decay score that combines recency, frequency, and intrinsic value with optimized weights:

• → Core: sufficient access frequency + high composite score + high importance
• → Peripheral: low composite score or extended inactivity
• Working: default tier for all new memories

Transition thresholds are derived from empirical testing, not hardcoded magic numbers.

Frequency Scoring

Raw access count can create a runaway advantage for frequently recalled memories, drowning out newer or less-accessed but relevant ones. Mnemo Cloud applies a soft logarithmic frequency cap — the first few accesses count at full value, then each subsequent doubling adds diminishing returns.

This keeps frequently recalled memories competitive without letting them dominate the ranking.

In Core (self-hosted), raw frequency count is used directly. Cloud applies the optimized frequency transform automatically.

Lifecycle Integration

The MemoryLifecycle class ties decay to tier transitions:

• A composite decay score combining recency, frequency, and intrinsic value is computed for each memory
• When the composite score crosses tier thresholds, the memory is promoted or demoted
• Stale peripheral memories below a minimum composite threshold are archived to JSONL files, keeping the active vector store lean
• There are no hard capacity limits — the decay model ensures irrelevant memories fade naturally through tier demotion and eventual archival

Configuration

Core provides configurable decay parameters. Sensible defaults are built in — most users don't need to change them:

const mnemo = await createMnemo({
  embedding: { /* ... */ },
  dbPath: './db',
  decay: {
    recencyHalfLifeDays: 30,  // how fast memories fade
  },
});

Mnemo Cloud uses optimized parameters tuned through extensive benchmarking. No configuration needed.