Concepts · Humans and machines
ACT-R-style agent memory uses goals, active context, retrieved chunks, and production rules to help AI agents choose what to do next.
Review pending. Report issues to editor@achiral.ai.
ACT-R Memory for Agents

ACT-R-style agent memory uses goals, active context, retrieved knowledge, and conditional action rules to guide an agent's next step. ACT-R is a cognitive architecture for modeling human cognition; applying its ideas to an AI workflow requires an explicit design for what to remember, when to retrieve it, and how to act on it.
The preceding ACT-R chapter distinguished evidence for content learning from examples for decision learning. Here we follow those roles through a longer task: a security agent investigating a new vulnerability advisory. This is a hypothetical workflow inspired by ACT-R, not an executable cognitive model or a description of Achiral's deployed implementation.
Give remembered knowledge and action rules different jobs
ACT-R's declarative memory holds chunks, structured units of knowledge. Buffers expose information from modules to current processing. Productions are condition–action rules that test buffer contents and can change them or request further processing. Kurup and colleagues describe these mechanisms in Using Expectations to Drive Cognitive Behavior.
In an agent workflow, these distinctions help separate three questions:
| Role | Question in the security case |
|---|---|
| Remembered knowledge | What package and service did the earlier scan identify? |
| Active task context | Which advisory and deployment are being checked now? |
| Conditional action rule | What should happen when the deployed version is still unknown? |
The Chunks and Production Rules chapters explain the representations. Naming a record a chunk does not supply the rules for using it.
Retrieve the earlier finding, then check the present
Suppose a new advisory affects a package that appeared in last month's scan of a service. The agent's current goal is to establish whether that service's deployed version falls within the advisory's affected range.
A retrieval request could use the service and package as cues. The old finding identifies the earlier version and its remediation record. It gives the investigation somewhere to start, but the service may have changed since the scan. An earlier patch may also address a different advisory.
The next action is therefore to inspect the current deployment, rather than close the finding from history alone. In a simplified workflow state:
Goal: assess this advisory for the target service
Remembered record: package appeared in an earlier scan
Current deployed version: unknown
Version check: not yet requested
Next step: request an authorized deployment-version checkThis is explanatory notation, not ACT-R syntax. A conditional rule would request the check only while the version is unknown and no check is pending. After the request is sent, the state becomes waiting; another pass through the workflow should not send the same request again.
If the check fails, the case remains unverified and needs a defined retry or review path. Missing evidence is not a negative finding.
Carry the decision through to an observed result
For this worked case, suppose the check reports a version within the affected range. The agent records that observation and opens a remediation ticket under the workflow's approval policy.
The handoff to a patching agent includes the advisory, target service, observed version, evidence source, required approval, and verification still to be done. That gives the recipient the basis for the task, rather than an unexplained instruction to patch something.
After approval, suppose the patching workflow deploys a fixed version and a fresh check confirms that version on the target service. The case can then record the observed version and mark this advisory's version check as resolved. That result does not establish that the service has no other vulnerabilities.
The record now contains both content and a decision sequence: an old finding prompted a current check; current evidence prompted remediation; confirmation justified the final update. Those are the distinct kinds of examples discussed in ACT-R and AI Memory.
Keep activation separate from action selection
In ACT-R, retrieval requests specify criteria for remembered chunks. Activation influences which matching chunk is retrieved and how quickly, subject to the retrieval threshold. Production selection uses matching buffer conditions and production utilities. The ACT-R account in Kurup and colleagues' paper distinguishes these mechanisms.
For the security workflow, retrieving a frequently used scan record would not make it current. Nor would high production utility establish that an action is permitted. The application must check evidence and authorization through its own rules.
Borrowing activation also creates a design question: what should happen when retrieval fails? This workflow could request a fresh inventory check without the old scan. Its ability to continue should not depend on recalling one particular historical record.
What Achiral takes from this design
At Achiral, the design lesson is to connect remembered evidence to a conditional next step. For this example, that means preserving the finding's scope, the current verification state, and the conditions behind the action.
This is architectural inspiration. It does not establish that Achiral runs an ACT-R model or implements its buffers and production selection. The memory documentation covers the product; this essay explains the design distinction.
To evaluate a workflow built on these ideas, include a stale scan, a missing deployment result, a check already pending, and an unconfirmed patch. Inspect whether each case produces an appropriate action and an accurate state update. A successful recall alone would not pass that evaluation.
Questions people ask
Is ACT-R the same as agent memory?
No. ACT-R is a cognitive architecture; agent memory describes software mechanisms for retaining and reusing information across work. An ACT-R-inspired design borrows selected ideas. Claiming to implement ACT-R requires specifying the actual cognitive model and mechanisms, not just storing records called chunks. See the ACT-R architecture introduction.
How is ACT-R-style memory different from RAG?
Retrieval-augmented generation uses retrieved information to condition generated output, as described in Lewis and colleagues' RAG paper. An ACT-R-inspired workflow also specifies goals, active state, and conditional actions. The approaches can coexist: RAG could supply advisory evidence while workflow rules determine whether to verify, wait, or escalate. See RAG vs ACT-R Memory for AI.
Does ACT-R make an AI agent human-like?
Borrowing ACT-R mechanisms does not establish human-like cognition or consciousness. A workflow must be evaluated on its task behavior. In this example, useful evidence would be correct version checks, appropriate handoffs, and accurate outcome records; none of those tests shows that the agent thinks like a person.
Why does activation matter for agents?
Activation provides a model of memory availability rather than treating every stored record as equally retrievable. In an agent design, it can inform how history and context influence recall. It does not measure a record's truth, freshness, or permission to act. The ACT-R memory chapter explains the underlying retrieval distinction.
Sources
- ACT-R research group, Carnegie Mellon University: the architecture and its research purpose.
- Kurup et al., “Using Expectations to Drive Cognitive Behavior”: chunks, buffers, retrieval activation, and production selection. The security workflow is our illustrative application, not a result reported in this paper.
- Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”: retrieval used to condition language generation.
Go back to ACT-R and AI Memory, continue with RAG vs ACT-R Memory for AI, or return to the Concepts hub.