Concepts · Humans and machines
How ACT-R represents knowledge as chunks, uses cues to request it, and models retrieval through activation, with a worked support-case example.
Review pending. Report issues to editor@achiral.ai.

Chunks and Declarative Memory
A chunk is a structured unit of knowledge in ACT-R, a cognitive architecture for modeling human thought and behavior. Declarative memory stores chunks that the model can retrieve using cues. A chunk represents information such as a fact or an event; its activation helps determine whether, and how quickly, it can be retrieved.
The previous chapter distinguished retaining information from being able to recover it. We now need a representation of that information and a way to request it.
Consider an assistant helping Susan reset her password. It has records of her earlier failed reset, a billing question, and a change to her notification settings. The immediate problem is to retrieve the failed reset without confusing it with another customer's case. This illustrative support workflow will let us separate three choices: what a chunk represents, which cues request it, and what makes retrieval succeed.
Give remembered information named fields
A chunk consists of named slots and their values. The ACT-R reference manual's section on chunks describes their representation. For the support example, we could represent the failed reset as:
user: Susan
task: password-reset
status: failed-unresolved
failure-code: unknownThese are illustrative fields, not ACT-R source code. The modeler chooses which details to represent. This record preserves a few facts about the failure, but omits Susan's words and the sequence of steps that preceded it. A later decision that depends on those omitted details would need another source of information.
Chunk types can declare slots and defaults, which helps organize a model. Current ACT-R also permits chunks without a declared type; a type is not an essential part of every chunk. That distinction is documented in the same manual.
Chunks also carry temporary information, including goals and perceived objects. A chunk's presence in a buffer does not, by itself, mean it is a durable record in declarative memory.
Use cues to specify what should come back
A retrieval request describes the information the model wants. In our example, its cues might be:
user: Susan
task: password-resetWith exact matching, both conditions must hold. Susan's billing record would fail the task condition. Another customer's password-reset record would fail the user condition. Susan's notification settings would fail the task condition too.
These conditions narrow the candidates, but they do not uniquely identify an event. If Susan has several password-reset records, more than one could match. A request that needs her latest unresolved failure must represent and handle those distinctions; the two cues above do not express them.
ACT-R can also use partial matching, allowing mismatches to incur penalties. Anderson and Betz discuss this mechanism in their model of categorization. For an AI product, a customer identifier would need an explicit identity check. Similarity to another customer's record would not establish permission to use it.
Activation affects retrieval success and timing
Matching a request does not guarantee retrieval. ACT-R uses activation to model the availability of declarative memories. Gunzelmann and Gluck describe its components and effects in their study of fatigue and cognitive performance.
A compact expression for these contributions is:
Here, is the activation used to evaluate chunk :
- is base-level activation, reflecting the recency and frequency of use.
- is spreading activation from the current context, through associations with other chunks.
- is a nonnegative mismatch penalty when partial matching is enabled.
- is activation noise, which introduces variation between retrievals.
The expression summarizes these mechanisms; their contribution depends on the model's settings. Among eligible candidates, the chunk with the highest activation can be retrieved if it exceeds the retrieval threshold. Higher activation also predicts shorter retrieval time. Below threshold, the request can fail even though a matching chunk is stored.
In Susan's case, cues establish which records match the request. Activation then helps determine which eligible record becomes available. A recent failed reset may be easier to retrieve, but recency alone does not make it the correct record.
Put the retrieved chunk to work
A successful retrieval places a chunk in the retrieval buffer, where production rules can test its contents. The reference manual describes this interface between declarative and procedural memory.
For Susan, the returned chunk could establish that her reset is unresolved and the failure code is unknown. Choosing whether to ask for the code, wait for an answer, or take another step requires an action rule. The next chapter will build that rule and track how it changes the case state.
A retrieval failure needs a response too. In the illustrative workflow, the assistant could ask Susan what happened or consult the case history. Treating a failed retrieval as evidence that no earlier failure occurred would confuse the availability of a memory with the history of the case.
Check the record before relying on it
A useful memory representation makes its contents inspectable and gives a retrieval request something precise to ask for. That lets a builder examine what information became available before asking how it influenced an action.
Activation is a retrieval mechanism, not a measure of truth or authorization. A frequently used record can remain wrong. An accurate record can become outdated. A record about Susan can still be unavailable to a particular agent under the product's access rules.
For this support case, the builder needs to check the retrieved record's identity and current status before using it. If a later record says Susan regained access, yesterday's failure remains part of the history but should no longer describe her present situation. Continue with Production Rules and Procedural Memory to turn the retrieved chunk into conditions for the next action.
Sources
- ACT-R reference manual: chunk representation, chunk types, buffers, and declarative retrieval.
- Anderson and Betz, “A hybrid model of categorization”: partial matching and retrieval thresholds.
- Gunzelmann and Gluck, “An integrative approach to understanding and predicting the consequences of fatigue on cognitive performance”: activation components and their effects on retrieval success and timing.
Go back to Ebbinghaus and Forgetting, continue with Production Rules and Procedural Memory, or return to the Concepts hub.