Short answer: They fade relevance on a curve where early time hurts more than later, often described with a memory half-life.
Linear decline assumes equal loss each week, which mismatches real usefulness. Exponential decay drops faster early then slows. Different categories need different half-lives for transient vs durable facts. Decay changes ranking weight, not the stored text itself. Weaviate supports curved recency effects on ranking.
The previous chapter treated recency as one signal among several feeding into a retention decision, weighted alongside importance and access frequency. This chapter looks more closely at recency’s own internal shape, the specific mathematical curve that describes how a memory’s relevance should actually fade over time, and why an exponential decay model tends to capture that fading more faithfully than a simple linear one.
Why Isn’t a Simple, Linear Decline in Relevance an Accurate Way to Model How Memories Actually Lose Their Standing Over Time?
A linear model assumes a memory loses the same fixed amount of relevance for every unit of time that passes, a week old costs exactly as much standing as a second week old, and a year old costs exactly as much as any other year. This doesn’t actually match how most information’s usefulness fades in practice, a memory usually loses a meaningful share of its relevance relatively quickly after it stops being actively useful, but whatever relevance survives that initial drop tends to persist much longer and fade far more slowly afterward. A straight line simply doesn’t capture that shape.
What Does an Exponential Decay Model Actually Describe That a Linear One Misses?
An exponential decay model describes relevance fading proportionally to however much relevance currently remains, rather than by a fixed amount per unit of time. This produces exactly the curve that matches real-world fading more closely, a steep initial drop while a memory is fresh and its specific details are still being superseded by newer information, followed by a much gentler, longer tail where whatever core fact survived that initial period keeps its remaining relevance for a long time afterward, rather than continuing to bleed away at the same rate it did at the very start.
What Does a Memory’s “Half-Life” Actually Mean in the Context of This Kind of Decay Model?
A half-life is the amount of time it takes for a memory’s relevance to fall to half of whatever value it started at, and it’s a genuinely useful way to reason about decay because it translates an abstract mathematical curve into something intuitively meaningful, how long does this kind of information stay roughly as useful as it started out being. A short half-life fits information that’s inherently short-lived, a temporary status update, a one-off scheduling detail, while a long half-life fits information that’s naturally durable, a stated long-term preference, a foundational fact about someone’s identity or role.
Why Does It Matter That Different Categories of Memory Might Genuinely Deserve Different Decay Rates Rather Than One Universal Curve Applied to Everything?
Treating every memory with the same decay rate would force a system to choose between a curve that’s too aggressive for durable facts, prematurely discounting information that’s still genuinely useful, or a curve that’s too gentle for genuinely short-lived information, letting stale, temporary details linger with more apparent relevance than they’ve actually earned. A system that lets different categories of memory decay at their own appropriate rates, a fast half-life for transient context and a slow one for durable facts, respects the genuine difference between these kinds of information rather than forcing an artificial uniformity onto data that was never actually uniform to begin with.
Does Applying a Decay Model Mean a Memory’s Content Actually Changes as Time Passes, or Only How It’s Weighted During Retrieval?
Only the weighting changes, not the underlying content itself, which is an important distinction to keep clear. Decay describes how much a memory’s age should count against it when it’s being ranked or considered for retention, it doesn’t rewrite or corrupt the memory’s actual stored facts. A decayed memory that still turns out to be the best available answer to a specific query can still surface and be used, its decay simply means it needs a stronger underlying match to overcome its reduced standing, rather than being automatically excluded the moment its age crosses some threshold.
How Does Weaviate Support Applying a Genuinely Curved, Rather Than Simply Linear, Decay to How Recency Affects a Memory’s Ranking?
Weaviate’s time-decay boost applies a curve, rather than a flat linear penalty, to how far a result sits from a reference date, letting a system tune how sharply relevance falls off with age to match the actual half-life a specific category of memory deserves. Consider a seasonal retail chain’s inventory-planning assistant, where a note about last month’s regional demand spike should fade in influence relatively quickly once the season passes, while a note about a store’s long-term customer demographics should barely fade at all over the same stretch of time:
import weaviate
from weaviate.classes.query import Boost
collection = client.collections.get("EngramMemories")
results = collection.query.hybrid(
query="What recent demand patterns should inform next month's inventory order?",
boost=Boost.time_decay(
weight=0.6,
),
limit=5,
)
Applying a stronger decay weight to this specific inventory-planning search means last season’s demand spike naturally loses influence as the calendar moves on, its steep initial relevance settling into a much gentler tail rather than either disappearing abruptly or lingering at full strength indefinitely, while the store’s separate, far more stable customer-demographics memory can be searched under a much lighter decay setting suited to information that’s genuinely durable. This is exactly the value a genuinely curved decay model delivers for a use case like seasonal retail planning, where a demand pattern’s real relevance falls off quickly at first and then levels out, a shape a simple linear penalty could never quite capture correctly.
Exponential decay gives a system a mathematically grounded way to let a memory’s relevance fade at whatever pace actually matches the kind of information it represents, steep and quick for the transient, gentle and long for the durable, rather than forcing every memory through the same uniform, linear treatment. This closes out the foundational discussion of forgetting, retention signals, and decay covered across these opening chapters of exploring how a memory system stays useful over time rather than accumulating indefinitely. The deliberate mechanisms surveyed here, importance, recency, frequency, and now decay curves, together form the toolkit a well-maintained memory system draws on to keep what matters and let go of what doesn’t, gracefully and on purpose rather than by accident. Our next chapter, What is soft forgetting in agent memory?, takes up accessibility decay that leaves memories recoverable.