A model from 1968
In 1968, Richard Atkinson and Richard Shiffrin published a paper that would define how psychologists think about human memory for the next half century. Their multi-store model proposed that memory is not a single system but a pipeline of three distinct stores, each with different capacities, durations, and encoding mechanisms.
The model was elegant in its simplicity: information enters through a sensory register, gets promoted to short-term memory if it receives attention, and eventually consolidates into long-term memory through rehearsal and repetition. At each stage, information can be lost. That loss is not a failure — it is the system working as designed.
When we set out to build a memory system for AI agents, we did not start with database schemas or vector embeddings. We started with this model. And it mapped perfectly.
The three stores
Sensory Register: everything the agent sees
In the human model, the sensory register captures raw perceptual input for fractions of a second. Most of it is discarded immediately. Only what you pay attention to moves forward.
In NEXO Brain, the sensory register is the session buffer. Every conversation produces a stream of information: user statements, agent responses, decisions made, errors encountered. The session buffer captures all of it as structured events. At session end, an extraction process identifies what matters — corrections, factual claims, preferences, decisions — and promotes those items to short-term memory. The rest is discarded.
This is the first layer of filtering. Without it, memory fills up with noise: greetings, thinking-out-loud, and context that was relevant only in the moment.
Short-Term Memory: the working set
Atkinson and Shiffrin described short-term memory as having limited capacity (roughly 7 items, per George Miller's famous paper) and rapid decay. Information stays here only if it is actively rehearsed.
NEXO's STM holds memories with a 7-day half-life. A memory created today starts at full strength (1.0). After 7 days without access, its strength drops to 0.5. After 14 days, 0.25. Eventually it falls below the retrieval threshold and becomes dormant.
But here is the key: every time a memory is retrieved, its strength resets. A fact you look up every day never decays. A decision you made once and never reference fades naturally. This mirrors how human short-term memory works: rehearsal keeps information alive, neglect lets it go.
Long-Term Memory: consolidated knowledge
Long-term memory in the Atkinson-Shiffrin model has theoretically unlimited capacity and much slower decay. Getting information there requires effort — repetition, meaningful encoding, or strong emotional association.
In NEXO, memories are promoted to LTM when they demonstrate sustained importance: repeated access, high relevance scores, or explicit pinning by the user. LTM memories have a 60-day half-life, giving them roughly 8x the persistence of STM. They are the agent's durable knowledge base: your deployment process, your team members' names, the architectural decisions behind your codebase.
Why forgetting is a feature
The Atkinson-Shiffrin model is as much about forgetting as it is about remembering. Information is lost at every stage. This was not a limitation the researchers lamented — it was a design principle they documented.
The same logic applies to AI agents. An agent that remembers everything is not smarter — it is slower and noisier. Every query returns more results, many of them outdated. The context window fills with stale information. Decision quality degrades because the signal-to-noise ratio drops.
NEXO implements forgetting through Ebbinghaus forgetting curves, the mathematical model Hermann Ebbinghaus published in 1885 describing how memory strength decays exponentially over time. The formula is straightforward:
strength = e^(-t / half_life)Where t is time since last access and half_life depends on the memory store (7 days for STM, 60 for LTM). Pinned memories are exempt from decay entirely.
Rehearsal and reinforcement
In human cognition, rehearsal is the mechanism that moves information from short-term to long-term memory. Repeating a phone number keeps it in STM. Understanding why the number matters — because it belongs to your doctor — encodes it into LTM.
NEXO models rehearsal in two ways. Access reinforcement resets the decay clock every time a memory is retrieved. Spreading activation strengthens connections between memories that are frequently retrieved together, building an associative network that mirrors how human knowledge is structured.
In practice, this means the agent naturally develops clusters of related knowledge. Memories about your deployment process link to memories about your staging environment, which link to memories about your team's code review workflow. Retrieving any node in the cluster activates its neighbors, producing richer, more contextual responses.
From theory to production
We have been running this architecture in production for months. Here is what we have observed:
- Memory stays lean. After 100+ sessions, the active memory set is typically a few hundred items, not thousands. Decay handles the cleanup automatically.
- Retrieval is fast and relevant. The combination of semantic search, recency weighting, and spreading activation consistently surfaces the right memories. On the LoCoMo benchmark, this architecture scores 72.1%.
- The "oh wait" moment. The most satisfying behavior: an agent surfaces a memory from weeks ago that is suddenly relevant. "You mentioned this was a recurring issue on March 12th." That is long-term memory working as designed.
- Corrections stick. When you correct the agent, that correction becomes a high-priority memory with a natural decay path. It influences behavior strongly at first and fades only if never reinforced.
Atkinson and Shiffrin could not have anticipated that their model would be used to build memory for artificial agents. But the principles they identified — attention gating, limited capacity, rehearsal, and decay — turn out to be exactly what AI agents need. The problem of managing knowledge over time is not new. Cognitive science solved it decades ago. We just had to listen.