Three-Store Model: The Blueprint Behind Every AI Memory Architecture

These

Atkinson and Shiffrin modeled human memory in 1968 as a linear processing chain: Sensory Register → Short-Term Memory → Long-Term Memory. The model has been largely superseded as psychology since the 1970s — Baddeley, Craik, and others demonstrated that memory operates neither serially nor passively. But as an architectural template, it is undead: context windows, vector stores, retrieval-augmented generation — the basic structure of today’s AI memory systems follows exactly this pipeline. This raises an uncomfortable question: if the underlying model makes the wrong assumptions about memory, which errors do the systems built on it inherit?

The Serial Pipeline

The original model proposes three stages that information traverses sequentially. The first stage — the sensory register — holds sensory impressions for under one second with unlimited capacity. The middle stage — short-term memory — maintains information for roughly 20 seconds, holding 5–9 units. The final stage — long-term memory — offers unlimited capacity with potentially permanent retention. Information that does not receive active processing decays at each stage.

Crucial is the distinction between structure and control processes: rehearsal maintains information in working memory, encoding transfers it to long-term storage through elaboration, retrieval accesses stored knowledge intentionally. The differentiation between simple repetition and elaborative rehearsal — linking new information to existing knowledge — remains relevant for designing contemporary memory architectures, because only the latter reliably produces long-term storage.

The Computer Metaphor Returns

The model emerged during the cognitive revolution of the 1960s, replacing behaviorism with information-processing frameworks. Atkinson and Shiffrin adopted the architecture of computer systems — input buffer, RAM, hard drive — and projected it onto brain function. This was a deliberately reductionist step that proved productive: it made memory modelable and experimentally accessible.

This concept transfer has since reversed direction. Artificial intelligence projects memory models back onto software — using the same three-tier structure that Atkinson and Shiffrin had borrowed from computers in the first place.

Where the Model Fails

Fundamental criticism emerged from the 1970s onward:

• Baddeley & Hitch (1974) replaced passive short-term memory with an active working memory model featuring phonological loop, visuospatial sketchpad, and central executive
• Craik & Lockhart (1972) argued that processing depth, not storage duration, determines retention
• Implicit learning bypasses the pipeline entirely — information reaches long-term memory without conscious rehearsal
• Patients with short-term memory deficits can still form new long-term memories, directly contradicting the serial assumption

These objections are not marginal. They concern the core assumption of the model: that information must pass through a linear chain in order to be stored durably.

The Architecture in AI Systems

Though outdated as psychology, the three-store architecture persists in modern AI systems:

Atkinson/Shiffrin (1968)	AI Agent Memory (2024–26)
Sensory Register	Input Buffer, Prompt Tokenization
Short-Term Memory	Context Window, Working Memory
Long-Term Memory	Vector Stores, Knowledge Bases, Persistent Memory
Rehearsal	Retrieval-Augmented Generation, Memory Consolidation
Encoding	Embedding, Chunking, Summarization
Decay/Forgetting	Decay Functions, TTL, Eviction Policies

The CoALA taxonomy (Cognitive Architectures for Language Agents, Princeton, 2023) — a framework defining four memory types for AI agents: Working, Procedural, Semantic, and Episodic Memory — refines this structure with Baddeley’s corrections. But the basic idea — that an agent needs different storage systems with different capacities and access times — comes directly from Atkinson/Shiffrin.

The Undervalued Sensory Register

AI memory discussions typically skip the sensory register — yet it solves a real problem: filtering before processing. Visual memory (iconic memory) holds impressions for roughly 300–500 milliseconds; auditory memory (echoic memory) persists for around 10 seconds. Both represent modality-specific pre-filters that separate relevant from irrelevant input before the limited capacity of short-term memory is engaged.

In AI systems, this corresponds to the step between raw input and tokenized prompt. Most current architectures lack explicit filtering at this level — everything proceeds directly into context windows. For multimodal agents processing text, images, and audio simultaneously, this becomes a bottleneck.

Einordnung

This text reads a psychological model as an architectural template for AI memory — from the perspective of a practitioner who builds knowledge systems and takes the conceptual precursors seriously. What this perspective can see: structural parallels between cognitive science models and software architecture, practical implications for memory design. What it cannot see: the empirical differentiation of the model by experimental cognitive psychology, the neurobiological foundations of memory processes, and the philosophical debate about the limits of the computer metaphor for mental phenomena. A cognitive psychologist would likely criticize the simplifications in the analogy table more sharply.

Kritische Einordnung

What Holds Up

• Multiple storage systems with different capacities and retention periods remain empirically robust
• The distinction between control processes and structure remains relevant for memory system design
• As historical foundation for AI memory research, the model is indispensable — CoALA, MemGPT, and SOAR all build on this tradition
• The computer metaphor returns as AI memory architecture — a productive circle demonstrating structural viability despite flawed details

What Needs Context

• Serial pipeline does not match reality: Human memory works in parallel, recursively, context-dependently — not as linear chains. AI systems that take the model literally reproduce this flaw
• Wikipedia as source: Encyclopedic summaries necessarily simplify and only roughly represent current research
• Analogy trap: Structural similarity suggests biological and digital memory solve identical problems — but biological memory reconstructs itself during retrieval, while digital storage reproduces. Confusing these produces systems optimizing the wrong target
• Missing emotional component: The original model ignored emotion’s effect on encoding. Contemporary AI memory systems share this blindness — lacking relevance signals beyond recency and similarity, though emotion is the strongest natural encoding amplifier

Diskussionsfragen

01 Architectural Legacy: If the three-store model implicitly shapes AI memory — which original flaws do current systems reproduce? Does the serial assumption (Input → Working Memory → Long-Term) remain embedded in today’s pipelines?

02 Sensory Register for AI: Do AI agents need explicit equivalents to sensory registers — modality-specific pre-filters before context windows? What are the implications for multimodal agents processing text, images, and audio simultaneously?

03 Forgetting as Design Principle: The original modeled decay as passive loss. Contemporary systems implement forgetting through active policies (decay functions, eviction). Which approach better serves knowledge OS architectures?

04 Levels of Processing Instead of Storage: Craik and Lockhart argued processing depth — not storage location — determines retention. What would this mean for AI memory design — shifting focus from storage locations to encoding processing depth?

Quellen

• Atkinson, R.C. & Shiffrin, R.M. (1968): Human Memory — A Proposed System and its Control Processes
• Wikipedia: Drei-Speicher-Modell
• Baddeley, A.D. & Hitch, G. (1974): Working Memory
• Craik, F.I.M. & Lockhart, R.S. (1972): Levels of Processing — A Framework for Memory Research
• Sumers et al. — CoALA: Cognitive Architectures for Language Agents (Princeton, 2023)
• Packer et al. — MemGPT: Towards LLMs as Operating Systems (2023)

Glossar

Three-Store Model (Atkinson-Shiffrin Model) 1968 model proposing three serial memory stages: sensory register, short-term memory, long-term memory. Historically influential; psychologically outdated.

Sensory Register Initial memory stage holding sensory impressions for fractions of a second. Unlimited capacity, extremely brief retention. Filters by relevance before information reaches working memory.

Rehearsal Control process maintaining information in working memory. Maintenance rehearsal involves simple repetition; elaborative rehearsal links information to existing knowledge — only the latter reliably produces long-term storage.

Levels of Processing Alternative model arguing processing depth, not storage location, determines retention. Deep processing produces more stable memories than shallow approaches.

Working Memory Model 1974 advancement replacing passive short-term memory with an active system featuring phonological loop, visuospatial sketchpad, central executive, and episodic buffer.

CoALA 2023 Princeton framework defining four memory types for AI agents — Working, Procedural, Semantic, Episodic Memory — continuing the cognitive science tradition beginning with Atkinson/Shiffrin.