Memory Decay: AI Memory That Forgets Like Humans Do

Most AI memory systems treat all memories equally. Something mentioned two years ago carries the same weight as yesterday's conversation. That's not how human memory works — and it creates awkward, irrelevant AI responses.

Today we launched Memory Decay, a feature that makes AI memory behave more like human memory. Frequently used memories stay strong. Unused ones naturally fade. The result is more relevant, contextual AI interactions.

How It Works

Memory Decay is modeled after the Ebbinghaus forgetting curve from cognitive psychology. Memories naturally decay over time unless reinforced through recall — exactly how your brain prioritizes what matters.

Confidence Scoring — When the AI extracts a memory, it assigns a certainty level. "User is vegetarian" (explicit statement) ranks higher than "User might prefer mornings" (inference). Higher confidence memories surface first.

Reinforcement — Every time a memory is recalled and used, it gets stronger. The decay timer resets and the memory receives a ranking boost. Frequently relevant memories survive. Unused ones fade.

Temporal Awareness — Time-sensitive memories like "traveling to Tokyo next week" decay faster than permanent facts like "favorite color is blue." Outdated information doesn't clutter your AI's context.

Source Tracking — Every memory tracks its origin: explicit (created via API), inferred (extracted from conversation), or corrected (user feedback). You always know where knowledge came from.

Why It Matters

• More relevant responses — Active, reinforced memories surface first

• Less noise — Outdated information naturally fades away

• Better context windows — Limited token budgets are spent on what matters

• Human-like interactions — AI that remembers the way people do

Pricing

Memory Decay is included free in all plans. Free tier includes 1,000 memory actions and 10,000 API calls per month.

Get Started

Memory Decay works automatically on all projects. No API changes required. View decay information — confidence scores, reinforcement counts, temporal flags — in the memory detail view on your dashboard.