Recency Bias vs Importance Scoring in AI Memory

What Recency Bias Looks Like

Recency bias manifests when a retrieval system returns a memory stored yesterday over a memory validated hundreds of times over the past year, simply because the yesterday memory has a more recent timestamp. The newer memory might be a casual observation, an untested hypothesis, or an edge-case remark, while the older memory is a well-established, frequently retrieved, highly corroborated piece of core knowledge. Pure recency ranking promotes the new over the proven.

In a customer support system, recency bias looks like a one-off note from an agent ("I think the refund policy changed") outranking the official policy documentation that has been retrieved by dozens of agents. In a development assistant, it looks like an experimental configuration from a feature branch outranking the production configuration that has been accessed every day for six months. In a personal assistant, it looks like a casual remark from yesterday ("maybe I should switch to a standing desk") outranking a well-established preference stated and confirmed repeatedly ("I work from the home office").

Why Pure Recency is Tempting

Recency is a strong signal for relevance, which is why it is tempting to weight it heavily. In many real-world scenarios, the most recent information is in fact the most useful. Product features change, code gets refactored, policies get updated, and preferences evolve. A system that always retrieves the most recent version of information is correct more often than one that retrieves a random version.

The problem is that "most recent" and "most reliable" are correlated but not identical. Recency works as a proxy for reliability when information updates are deliberate, reviewed, and authoritative (like documentation updates). It fails as a proxy when information accumulates organically from conversations, logs, and observations where any individual entry might be preliminary, speculative, or simply wrong.

What Importance Scoring Captures

Importance scoring measures how much the system actually relies on a memory, independent of when it was created. Three factors contribute to importance:

Access frequency: A memory retrieved 50 times is more important than one retrieved once, because usage validates utility. The system's users have implicitly confirmed the memory's value by repeatedly finding it useful. This is a stronger signal than creation timestamp because it reflects demonstrated utility rather than assumed freshness.

Corroboration: A memory supported by three independent sources is more important than one from a single source. The consolidation process detects when multiple memories support the same claim and increases the confidence of the corroborated memory. High corroboration means the information has been validated from multiple angles, making it more reliable regardless of age.

Entity centrality: A memory that connects to many entities in the knowledge graph is more important than one connected to few or none. High-centrality memories serve as knowledge hubs that multiple retrieval paths pass through. They represent foundational knowledge that ties different parts of the domain together.

The Balance Problem

Neither pure recency nor pure importance produces optimal rankings. Pure recency ignores established knowledge and chases novelty. Pure importance ignores change and becomes stale. The right approach blends both, with the blend ratio tuned to the domain's rate of change.

ACT-R's base-level activation equation provides an elegant solution because it naturally combines recency and frequency in a single value. Every access event contributes to activation, weighted by how recently it occurred. A memory accessed once yesterday has moderate activation from recency. A memory accessed 30 times over the past month has high activation from frequency, even though its most recent access might be a few days old. A memory accessed frequently for six months and then not at all for two months has declining but still significant activation from its deep access history.

This combined signal avoids both extremes. A brand-new, unvalidated memory does not automatically outrank established knowledge because it has only one access event. But an old memory that is no longer being accessed gradually yields to newer information that is actively being retrieved. The transition is gradual and proportional, not abrupt.

Domain-Specific Tuning

The right recency-importance balance varies dramatically across domains:

News and current events: Information becomes outdated within hours. Recency should dominate (80/20 recency to importance). Old news stories should rapidly fade from top results.

Customer support: Product changes invalidate old information within weeks. Moderate recency bias (60/40) keeps current product knowledge accessible while preserving frequently validated support patterns.

Software development: Code changes frequently but architectural knowledge persists for months or years. Balanced weighting (50/50) preserves foundational knowledge while surfacing recent code changes.

Legal and compliance: Regulations change slowly and old precedents remain relevant indefinitely. Importance should dominate (30/70 recency to importance) to ensure established rulings and regulations are not displaced by recent but less authoritative commentary.

Personal preferences: Core preferences are stable, but context changes. Moderate importance bias (40/60) protects deeply established preferences while allowing recent context to influence retrieval.

Confidence as a Stabilizer

Confidence scoring acts as a stabilizer between recency and importance. Memories with high confidence (above 8.0 in Adaptive Recall's model) receive partial protection from recency-driven decay. This means that even as a high-confidence memory ages and loses activation, its confidence score prevents it from falling too far in the rankings. The logic is that a fact confirmed by multiple sources and never contradicted is probably still true, regardless of how recently it was accessed.

This protection is not absolute. A high-confidence memory can still be displaced by a newer memory with both high recency and high similarity, because the newer memory is more likely to reflect current state. But it prevents the common failure mode where casual, unverified recent observations crowd out established, validated knowledge in the top results.

The confidence mechanism also handles the case where established knowledge is wrong. If a new memory explicitly contradicts a high-confidence old memory, the consolidation process detects the contradiction, which reduces the old memory's confidence. This makes it vulnerable to recency-driven displacement, which is the correct behavior: new evidence that contradicts old knowledge should eventually replace it.