Short-Term to Long-Term Memory Promotion Explained

The Short-Term and Long-Term Distinction

In cognitive science, short-term memory and long-term memory are distinct systems with different capacities, durations, and encoding mechanisms. Short-term memory holds a small number of items (classically seven, plus or minus two) for seconds to minutes. Long-term memory holds a practically unlimited number of items for days to decades. Transfer from short-term to long-term requires encoding processes like rehearsal, elaboration, and consolidation.

AI memory systems do not have the same biological constraints, but the distinction between provisional and established knowledge is equally important. A memory stored five minutes ago from a single conversational observation is provisional. It might be correct, it might be a misunderstanding, it might be out of date by tomorrow. A memory that has been retrieved twenty times over three months, corroborated by five independent sources, and connected to a dozen other memories through shared entities is established. Treating both with equal authority is a design error that degrades retrieval quality.

The promotion mechanism creates this distinction automatically. New memories enter with moderate activation and confidence. If they prove useful through retrieval and accurate through corroboration, they accumulate activation and confidence, becoming increasingly prominent in results and resistant to decay. If they sit untouched, their activation declines, and they gradually fade from retrieval results without ever reaching the established status that would protect them from forgetting.

How Promotion Works in Practice

Promotion in Adaptive Recall operates through three channels: access-based activation, confidence accumulation, and graph integration.

Access-based activation follows ACT-R's base-level learning equation. Every retrieval event adds a new term to the activation sum, weighted by recency through power-law decay. A memory retrieved once has a single activation contribution that decays over time. A memory retrieved five times has five contributions, each decaying independently. The cumulative effect is that frequently retrieved memories maintain high activation even as individual access events age. This is the primary promotion mechanism because it directly reflects demonstrated utility.

Confidence accumulation operates through corroboration. When the system detects that a new memory supports an existing one, the existing memory's confidence increases. If a developer stores a memory about "the team uses PostgreSQL for the main database" and three subsequent conversations reference PostgreSQL in the same context, the original memory's confidence climbs. High confidence (above 8.0 in Adaptive Recall) marks a memory as protected, which means it resists decay even during periods of low direct access. Confidence is the mechanism that promotes factual knowledge to long-term status even if it is not retrieved frequently, because some facts are foundational and do not need to be actively queried to remain important.

Graph integration measures how connected a memory becomes in the knowledge graph over time. When a new memory is stored, it connects to entities extracted from its content. As more related memories are stored, the original memory gains connections through shared entities. A memory about "PostgreSQL configuration" that was initially connected to three entities might be connected to twenty after months of related memories are added. This increasing connectivity is a form of promotion because it makes the memory reachable through more spreading activation pathways, which means it contributes to retrieval results for a wider range of queries.

The Probation Period

New memories effectively go through a probation period where they must prove their value before earning long-term status. This period is not configured explicitly but emerges from the interaction of moderate initial activation, the decay function, and the promotion mechanisms described above.

In practical terms, a new memory has about two to four weeks of moderate activation before decay reduces it to the point where it needs reinforcement to stay prominent. During this window, the memory is included in retrieval results when relevant queries arrive. If it is retrieved and proves useful, it gains activation that extends its effective lifespan. If it is corroborated by other sources, its confidence increases, providing additional protection against decay. If neither happens, the memory quietly fades from the top retrieval results and eventually falls below the forgetting threshold.

This probation period is a natural quality filter. Memories that encode useful, accurate knowledge tend to get retrieved and corroborated because they match real queries and align with other stored knowledge. Memories that encode noise, misunderstandings, or one-time observations that are never relevant again simply fade away. No manual review is needed because the system learns what matters from its own usage patterns.

Spaced Repetition and Natural Promotion

The spaced repetition effect, well-known in education and flashcard systems, emerges naturally from ACT-R's activation model. A memory accessed at regular intervals builds stronger long-term activation than one accessed many times in a burst. This is because each spaced access occurs after some decay has happened, and the fact that the memory was still retrievable after a decay period strengthens its long-term activation more than an access that happens while activation is already high.

In AI memory systems, this means that memories used consistently across sessions are promoted more effectively than memories used heavily in a single session. A coding pattern referenced once per week for two months builds very strong activation, stronger than the same number of accesses concentrated in a single day. This naturally promotes the knowledge that developers use as part of their regular workflow and allows session-specific details to decay appropriately.

Comparing Promotion Strategies

Some memory systems use explicit promotion rules instead of activation-based promotion. For example, a system might promote a memory to long-term status after it has been accessed three times, or after it has existed for 30 days without being contradicted. These rule-based approaches are simpler to implement but less nuanced than activation-based promotion.

The problem with threshold rules is that they treat all accesses equally. Three accesses in a single session are very different from three accesses spread over three weeks, but a simple count-based rule cannot distinguish them. Similarly, surviving 30 days without contradiction could mean the memory is accurate and well-established, or it could mean the system simply has not encountered the contradicting information yet.

Activation-based promotion, as implemented in Adaptive Recall, handles these nuances because the mathematics inherently weight recency, frequency, and spacing. It is more complex to implement than simple rules, but it produces promotion decisions that better reflect the actual quality and utility of each memory.