ACT-R (Adaptive Control of Thought - Rational), developed by John Anderson (1983, 1993, 2007), is a comprehensive cognitive architecture in which declarative memory plays a central role. The declarative module stores factual knowledge as "chunks" that are retrieved based on their activation levels, with activation determined by a principled combination of past usage history, current context, and stochastic noise.
Base-Level Activation
The base-level activation of a chunk reflects how recently and frequently it has been used. It follows a power-law decay function that is rationally adapted to the statistics of the environment (Anderson & Schooler, 1991):
where the sum is over all n presentations of chunk i, tⱼ is the time since the j-th presentation, and d is the decay parameter (typically d ≈ 0.5). This equation captures both the frequency effect (more presentations yield higher activation) and the recency effect (recent presentations contribute more), with the power-law decay ensuring that old memories persist but become less accessible.
Spreading Activation
Chunks also receive activation from the current goal and context through spreading activation. Each element in the current goal sends activation to chunks that are associated with it:
where Wⱼ is the attentional weight on source j (constrained so that Σ Wⱼ = W, a total source activation), Sⱼᵢ is the strength of association from source j to chunk i (typically Sⱼᵢ = S − ln(fanⱼ), where fanⱼ is the number of chunks associated with source j), and ε is noise drawn from a logistic distribution. The fan effect (slower retrieval when a cue is associated with many items) follows naturally from the log-fan term in the associative strength.
Retrieval Probability and Latency
A chunk is retrieved if its activation exceeds a threshold τ. The probability of retrieval is a logistic (softmax) function of the difference between activation and threshold:
Retrieval latency follows an exponential function of activation: T = F · e^(−f·Aᵢ), where F and f are scaling parameters. Highly activated chunks are retrieved quickly and reliably, while weakly activated chunks are slow and error-prone. This produces the speed-accuracy correlation that is ubiquitous in memory retrieval.
Rational Analysis
A distinctive feature of ACT-R's memory system is its grounding in rational analysis (Anderson, 1990). The activation equation approximates the log-odds that a chunk will be needed given the current context and the history of use. The base-level equation reflects the prior probability (based on past usage), while spreading activation reflects the likelihood (given the current context). This Bayesian interpretation makes ACT-R's declarative memory not just a process model but a rational, optimal system adapted to the statistical structure of the environment.
ACT-R's declarative memory module has been used to model an extraordinary range of phenomena: the fan effect in sentence verification, the spacing effect in learning, memory for past actions, analogy, categorization, and skill acquisition. The architecture is implemented as a freely available software system that runs as a production-system simulation, making it one of the most widely used computational cognitive models in psychology and cognitive science.