Memory decay

You remember yesterday's meeting better than last month's. You recall a fact you use every day more easily than one you read once. Human memory works this way because it has to. Infinite recall with no prioritization would be as useless as forgetting everything.

Nowledge Mem applies a similar idea. Every memory carries two independent scores that influence where it appears in search results. One fades with time. The other only grows.

Two scores, two jobs

Decay score reflects freshness. A memory you saved today scores higher than one you saved six months ago, all else equal. The score drops over time following an exponential curve, with a frequency boost for memories you keep coming back to. If you access a memory regularly, the decay score stays high.

Confidence score reflects how well-supported a memory is. It starts at a baseline and increases as evidence accumulates. Confidence never decreases. A memory that has been accessed many times, linked through EVOLVES chains, or used as a source for a Crystal will have higher confidence than an isolated, untouched memory.

The two scores are independent. A memory can be old (low decay) but well-supported (high confidence), or recent (high decay) but unvalidated (low confidence). Both signals feed into the final search ranking alongside semantic relevance.

What feeds into each score

Decay score

Two components, blended:

• Recency: exponential decay from the last time you interacted with the memory. Recent memories score higher. The curve has a half-life measured in weeks, not hours, so memories do not vanish overnight.
• Frequency: a logarithmic function of how many times the memory has been accessed. Frequent use keeps a memory fresh even if the last access was not today.

Recency carries more weight than frequency. A memory you used heavily last year but have not touched since will still fade, just more slowly than one you never accessed.

Confidence score

Six signals, each capped so no single signal can dominate:

1. Access frequency — how often the memory has been retrieved
2. Search appearances — how often it shows up in search results
3. Explicit clicks — how often you open and read it
4. Reading time — how long you spend on it when you do open it
5. EVOLVES edges — how many other memories confirm or enrich it
6. Crystal membership — whether it has been used as a source for a Crystal

These signals are tracked automatically. You do not need to rate or tag memories for the system to learn which ones matter.

The importance floor

Decay creates a problem: a genuinely important fact that you do not access for months will score low, even though it is still correct and valuable. Spaced repetition systems solve this by scheduling reviews. Mem is not a flashcard app.

Instead, each memory has an importance floor. The decay score can drop, but it will never fall below a minimum threshold tied to the memory's importance level. A foundational decision or a critical procedure stays retrievable even if you have not looked at it in a long time.

The floor is modest. It does not override semantic relevance. It just prevents "ghost memories" — knowledge that exists in your system but effectively cannot be found.

How decay affects search ranking

Semantic relevance is the dominant factor in search results. Decay and confidence are secondary signals that adjust the ordering when semantic scores are close.

In practice:

• Two memories equally relevant to your query → the fresher one ranks higher
• A semantically strong match always beats a recent but weakly relevant one
• A well-supported memory (high confidence) gets a small additional boost
• Crystals and latest versions in EVOLVES chains get their own ranking adjustments on top of this

Automatic refresh

The system recalculates decay and confidence scores daily as a background task. This refresh also updates the cached scores used in search, so results reflect current usage patterns without any manual intervention.

Optionally, you can enable auto-archival. Memories that fall below a very low decay threshold, have never been accessed, and are old enough can be moved to an archived state. Auto-archival is off by default and requires explicit opt-in. Archived memories are not deleted. They can be restored at any time.

Next steps

• Search architecture explains how decay and confidence feed into the full ranking pipeline
• Knowledge evolution covers EVOLVES chains, which contribute to confidence scores
• Crystals explains Crystal membership, another confidence signal