The Four Stages of AI Memory: Create to Delete

Stage 1: Creation

A memory begins when the system stores a new piece of knowledge. In Adaptive Recall, the store tool handles creation, accepting content from the calling application and producing a complete memory record with several automatically generated components.

The content itself is the text of the memory, a fact, observation, preference, decision, or any other piece of knowledge worth preserving. From this content, the system generates a vector embedding for semantic search, extracts entities and relationships for the knowledge graph, assigns an initial confidence score (typically 5.0 on a 0-10 scale, reflecting uncorroborated but plausible information), and sets the creation timestamp that starts the decay clock.

At creation time, the memory also establishes its position in the knowledge graph. If the content mentions "Kubernetes deployment," the system extracts that entity and connects the new memory to every other memory that references Kubernetes or deployment. These connections are bidirectional, so existing memories also become connected to the new one. This is important because it means that storing new knowledge automatically enriches the retrieval pathways for existing knowledge through spreading activation.

New memories enter with moderate activation. They are accessible through retrieval but do not dominate results. Their influence grows or fades based on what happens in the next stages.

Stage 2: Promotion

Promotion is not a discrete event but a continuous process driven by usage. Every time a memory is retrieved and used, its base-level activation increases according to ACT-R's activation equations. Memories that prove useful, those that get retrieved frequently and recently, accumulate activation that makes them increasingly prominent in future retrieval results.

Promotion also operates through confidence. When a new memory corroborates an existing one, the existing memory's confidence score increases. If three separate conversations all reference the same architectural decision, the memory encoding that decision gains confidence with each corroboration. High confidence serves as evidence that the memory is accurate and well-established, not just recently stored.

The combination of activation and confidence creates a natural priority system. Memories that are both frequently used and well-corroborated rise to the top of retrieval rankings. Memories that were stored once and never retrieved or confirmed remain at moderate levels and eventually begin to decline. This mirrors how human expertise works: knowledge that you have applied successfully many times feels more certain and comes to mind more readily than something you read once and never used.

In practical terms, promotion means that the first week of a new memory's life is a probation period. If it gets retrieved and proves useful, its activation and confidence grow, securing its place in the store. If it sits untouched, it begins the slow decline toward forgetting. No manual intervention is needed because the usage patterns themselves determine which memories matter.

Stage 3: Consolidation

Over time, a memory store accumulates multiple entries about the same topic, captured at different times, in different contexts, with different levels of detail. A developer might store memories about configuring a CI/CD pipeline across five different sessions, each capturing a piece of the overall picture. Consolidation merges these fragments into a single, comprehensive memory that carries the combined activation and confidence of all its sources.

The consolidation process, triggered by Adaptive Recall's reflect tool, works in phases. First, it clusters related memories by analyzing entity overlap and semantic similarity. Memories that share multiple entities or have very similar content are grouped together. Second, it evaluates each cluster for redundancy, identifying memories that say essentially the same thing in different words. Third, it checks for contradictions where memories in the same cluster make conflicting factual claims. Fourth, it produces merged memories that incorporate the most complete, most recent, and most confident information from the cluster.

Consolidation is modeled on what the human brain does during sleep. Neuroscience research shows that during sleep, the brain replays recent experiences, strengthens important connections, weakens unimportant ones, and reorganizes knowledge into more efficient structures. The result is that you wake up with clearer, better-organized knowledge than you went to bed with. Memory consolidation in AI systems produces the same effect: after a consolidation run, the memory store is smaller, better organized, and higher quality than before.

The metadata of consolidated memories reflects their combined provenance. The activation value inherits the strongest activation from the group. The confidence score increases because merging corroborating sources is itself a form of evidence. The entity connections are the union of all source memories' connections, which means consolidated memories are reachable through more graph traversal paths than any individual source was.

Stage 4: Forgetting

Forgetting is the final stage, and it is the one that most AI memory systems skip entirely, to their detriment. Every memory that is not regularly accessed or explicitly protected gradually loses activation through power-law decay. When activation drops below a configurable threshold, the memory is removed from the active retrieval index.

Removal can take two forms. Archival moves the memory to cold storage where it can be restored if it becomes relevant again. This is appropriate for knowledge that might cycle back into relevance, such as seasonal patterns or project-specific information that could apply to a future similar project. Permanent deletion removes the memory entirely. This is appropriate for information that is definitively outdated, such as a superseded API version or a resolved bug report.

The rate of forgetting is not uniform across all memories. Importance scoring modifies the effective decay rate for each memory based on access frequency, confidence, and entity centrality. High-importance memories, those that have been used extensively, are well-corroborated, and connect to many other memories in the graph, decay at a fraction of the normal rate. They persist in the store for months or years even without new accesses. Low-importance memories decay at the full rate and reach the forgetting threshold within weeks.

Controlled forgetting keeps the memory store focused. Without it, every transient observation, every outdated fact, and every corrected mistake remains forever, competing with current knowledge for retrieval slots. A memory system that forgets appropriately is like an expert who stays current: they do not clutter their thinking with outdated information, and when you ask them a question, you get the answer that reflects the current state of the world.

The Stages in Practice

Consider a concrete example. A developer stores a memory: "The payment service uses Stripe API v2023-10-16." At creation, this memory gets moderate confidence and activation. Over the next few weeks, the memory is retrieved several times when questions about payment processing arise, and its activation grows through promotion. Two months later, the team upgrades to a newer Stripe API version, and a new memory is stored: "The payment service was migrated to Stripe API v2024-08-01." During the next consolidation run, the system detects the contradiction between the two API version claims, resolves it in favor of the newer memory (recency wins), and merges any compatible information from both. The outdated version reference is removed. The old memory, now superseded, begins decaying. Eventually it falls below the forgetting threshold and is archived.

At no point did anyone manually tag, expire, or delete the outdated memory. The lifecycle handled it automatically through the interaction of promotion, consolidation, and forgetting. This is the power of a managed memory lifecycle: the system maintains its own accuracy through the same mechanisms that keep human memory focused and current.