AchiralAchiral

Concepts · Humans and machines

RAG retrieves evidence for generation; AI memory lets prior experience change what becomes salient and how a system responds or acts next.

Published2026-10-09720 reads

Review pending. Report issues to editor@achiral.ai.

RAG vs ACT-R Memory for AI

Watercolor illustration of a stone wine cellar with one selected bottle on a tasting table, bottles aging on shelves, a thermometer, and an open notebook.
Selecting a bottle illustrates retrieval. Remembered tastings can change which bottle is chosen next; the notebook evokes the experience behind that choice. Keeping the bottles alone does not provide that adaptation.

RAG retrieves evidence to inform generation. AI memory, in Achiral’s canonical definition, is a persistent change caused by prior experience that alters a system’s future behavior, responses, or internal dynamics. The distinction is between finding information for a task and experience changing what influences subsequent work.

Retrieval can be part of a memory system. A RAG application can also include learning and state updates; the label alone does not specify those mechanisms. The preceding article followed remembered evidence through an agent workflow. Here we ask what changes through experience, beyond which source is retrieved.

A wine cellar is a useful picture. Retrieval finds a bottle for this meal. Remembering a previous tasting can change which bottle is considered next time. Maintaining the inventory supports that process, but the inventory alone does not explain the change in choice.

A terminology caveat

Retrieval-augmented generation (RAG) has a relatively specific meaning in machine learning: retrieve information from an external source and use it to condition a model's output. The source may be a document index, database, search engine, or knowledge graph. Lewis and colleagues’ RAG paper describes conditioning generation on retrieved passages.

AI memory is a less standardized term. Here we use Achiral’s behavioral definition. Our Emergent Memory framework describes formation, association, activation, reinforcement, revision, and decay as mechanisms through which accumulated experience shapes what enters working memory and influences action. Retention and access controls support that process; they are not its defining behavior.

Different design objectives, not competing categories

DimensionRAGPersistent AI memory
Primary questionWhat evidence would inform this response?What from prior experience should influence this decision?
Core operationRetrieve and condition generation.Form and revise memories; change their availability and influence through use.
Across interactionsCan retrieve context repeatedly across a continuing workflow.Experience changes subsequent salience, responses, or actions.
Main correctness concernDoes the evidence support the generated answer?Does prior experience influence the next task appropriately, with its scope and uncertainty preserved?
RelationshipCan retrieve from a memory system.May use RAG as one retrieval mechanism.

A concrete example

Consider a hypothetical assistant helping a support team. When asked about a refund, it can use RAG to retrieve the current refund policy and cite the relevant section. That is request-time grounding.

Suppose a manager approves a one-time refund exception for this customer’s current order. The system records the approval and associates it with the unresolved case. A memory-guided design could then make the exception salient when work on that case resumes, so the assistant checks its status before applying the standard policy. The prior event has changed which evidence receives attention and which step comes first.

If successful follow-ups reinforce that association, the relevant history may become easier to recall in the same context. If the manager withdraws the exception, revision must change its influence: the old approval can remain historical evidence without authorizing another refund. These are proposed behaviors for this example, not reported results from a deployed system.

Scope, expiry, and permissions still matter. They keep this order’s exception from becoming a general refund rule. But the behavioral question is what the experience changes when the case returns, not merely whether the approval remains stored.

RAG vs. ACT-R Memory for AI

ACT-R separates declarative knowledge, current processing context, and conditional action rules. Chunks represent knowledge; buffers expose information to current processing; productions test conditions and specify actions. Kurup and colleagues describe these mechanisms in Using Expectations to Drive Cognitive Behavior.

ACT-R’s activation mechanisms offer a model of how use history and context affect recall; productions supply conditional actions. Those distinctions help describe how experience could influence an agent’s next step. Borrowing them does not establish a full ACT-R implementation or human-like cognition. RAG can supply evidence within such a design without accounting for all of its memory dynamics.

Design implications

For the support workflow, inspect how the next decision changes after the exception is approved, reinforced through relevant use, and later withdrawn. Holding a copy of the approval is insufficient if it never affects the next step, or if it keeps affecting the wrong case. The memory documentation covers product behavior; this example explains the design criterion.

Evaluation should compare behavior before and after the relevant experience, then check whether correction changes it again. Test an unrelated order as well. Increased recall is useful only if the memory influences the appropriate task without spreading an exception beyond its conditions.

Questions people ask

Can RAG be part of an AI memory system?

Yes. RAG can retrieve evidence from a memory system and use it to inform generation. Memory dynamics can shape what becomes available for retrieval or current reasoning, and outcomes can feed back into those dynamics. The mechanisms can coexist in one system.

Does RAG remember previous conversations?

A RAG pipeline can retrieve earlier conversations if they have been stored and made searchable. That provides access to history. Whether experience changes subsequent salience or action depends on the surrounding memory and decision mechanisms, not the RAG label alone.

When does an agent need more than retrieval?

When accumulated experience should change what it notices, recalls, or does next. In the refund example, the earlier approval should affect how the same case is handled, and its withdrawal should revise that influence. Retrieval supplies information; additional mechanisms define how experience changes subsequent behavior.

Continue with All Kinds of Memory in AI & Software for the behavioral definition and the mechanisms that support it.


Sources