How do recency and access frequency guide retention?

The previous chapter looked at importance as one signal for deciding what a memory system should hold onto. This chapter looks at two more signals that matter independently of how important a memory seemed when it was first captured, how recently it was created or updated, and how often it’s actually been retrieved and used since. Neither signal requires guessing at significance at all, both come directly from a memory’s own observable history.

Why Does Recency Matter as a Retention Signal Even for a Memory That Seemed Genuinely Important When It Was First Captured?

Importance measures how much a fact mattered at the moment it was recorded, but it says nothing about whether that fact is still current now. A memory describing a customer’s active subscription tier, genuinely important when it was captured, can become misleading the moment that subscription changes, and nothing about its original importance score reflects that shift. Recency captures a different, complementary dimension entirely, how much time has actually passed since a memory was last confirmed or updated, which matters specifically because the world a memory describes keeps changing even when the memory itself doesn’t.

What Does Access Frequency Actually Reveal About a Memory That Recency and Importance Both Miss?

Access frequency measures how often a specific memory has actually been retrieved and used in practice, which is a genuinely different signal from how important it seemed at creation or how recently it was written. A memory that gets retrieved constantly is demonstrating its own ongoing usefulness empirically, through real, observed demand, rather than through a judgment made once and never revisited. A memory that’s never once been retrieved since it was created, regardless of how important it seemed when it was first captured, is quietly signaling that it may not actually be earning its place in active memory.

Why Do These Two Signals Actually Work Better Combined Than Either One Does Applied Alone?

A memory that’s both recent and frequently accessed is about as strong a candidate for continued retention as a signal-driven system can identify, current and actively demonstrating its own usefulness at the same time. A memory that’s old but still frequently accessed, a foundational fact that genuinely doesn’t change, deserves to survive despite its age, since frequency is telling a system something recency alone would have missed. Conversely, a memory that’s recent but has never once been accessed might simply not have had a chance to prove itself yet, and treating recency alone as sufficient would risk discarding something too early, before frequency ever had the opportunity to weigh in.

Does Relying on Recency as a Retention Signal Risk Unfairly Penalizing Genuinely Stable, Long-Term Facts That Simply Haven’t Needed to Change?

Yes, and this is exactly why recency should inform retention decisions rather than dictate them outright on its own. A fact about someone’s long-standing professional background, or an organization’s founding history, doesn’t become less true simply because it hasn’t been touched or updated recently, and treating raw age as an automatic signal of declining relevance would unfairly punish exactly the kind of durable, foundational information a system most wants to keep confidently available. This is precisely why access frequency matters as a genuine counterbalance, a stable fact that keeps getting retrieved despite its age is demonstrating real, ongoing relevance that a naive recency-only approach would have missed entirely.

How Should a System Actually Combine Recency and Frequency Into a Single, Usable Retention Signal Rather Than Treating Them as Two Separate, Unrelated Numbers?

A practical approach weights recent access more heavily than a distant one while still counting every retrieval as some evidence of ongoing usefulness, letting a memory that was heavily used months ago but has gone quiet since gradually lose standing relative to one that’s being actively retrieved right now. This produces a genuinely dynamic signal rather than a static one calculated once and left alone, a memory’s standing can rise again if it starts getting retrieved more often, and can decline gradually if genuine demand for it actually fades, rather than the system permanently locking in a judgment made at some single, arbitrary point in time.

How Does Weaviate Support Factoring Recency Directly Into How Memories Actually Get Ranked at Search Time?

Weaviate’s boost mechanism includes a time-decay option that re-scores retrieved candidates based on how far they sit from a reference date, letting recency shape a search’s final ranking without ever excluding an older result outright the way a hard filter would. Consider a corporate knowledge-management system tracking internal process documentation, where a newer procedure should generally surface ahead of an older, possibly superseded one, but a genuinely foundational policy that hasn’t changed in years shouldn’t be buried purely because of its age:

import weaviate
from weaviate.classes.query import Boost

collection = client.collections.get("EngramMemories")

results = collection.query.hybrid(
    query="What is our current process for approving vendor contracts?",
    boost=Boost.time_decay(
        weight=0.3,
    ),
    limit=5,
)

Applying a moderate recency weight lets a newer version of the vendor-approval process naturally rise above an older, superseded draft when both are semantically similar matches for the same query, while still leaving room for a genuinely foundational, rarely-updated policy to surface when it’s actually the strongest match, since the boost adjusts ranking rather than excluding anything outright. This is exactly the value factoring recency directly into ranking delivers for a use case like internal process documentation, where an employee searching for current guidance needs the newest applicable answer to naturally rise to the top, without a rigid age cutoff accidentally hiding a foundational policy that’s simply never needed to change.

Recency and access frequency give a system two concrete, observable signals for retention decisions, distinct from importance and genuinely complementary to it, one reflecting how current a memory still is and the other reflecting how much real, ongoing demand it’s actually earning. Recency in particular can be modeled more precisely than a simple weighted boost, treating a memory’s relevance as something that fades along a deliberate mathematical curve over time. Our next chapter, What are exponential decay models for memory relevance?, takes up exactly that precision.