An agent’s current context contains the information supplied for its next model response. Memory preserves selected information for later use. Retrieval finds information and brings relevant material into context. Something can be saved without being visible now, and visible now without being current or correct.
For digital and IT leaders, business owners, and anyone learning how agents work, the useful question is more specific than “Does it remember?” Ask: What information will this decision use, where did it come from, and what needs checking again?
Four ideas that explain the difference
The model also brings knowledge learned during training. Separating that knowledge from the surrounding application’s records helps explain why an agent can describe a business process yet still need to look up your organization’s current rules.
Model knowledge: what training supplied
Model parameters are the learned numerical values that shape its responses. Lewis and colleagues’ foundational retrieval-augmented generation paper distinguishes knowledge held in model parameters from an external information store. Their system combined a trained model with retrieved Wikipedia passages. RAG paper.
That distinction matters operationally: knowing how equipment approvals generally work does not establish which approval policy applies to your request. Saving a policy document in an application’s memory also does not, by itself, retrain the model.
Context: what is available for this response
Context is the material presented to the model for a particular response, potentially including instructions, conversation history, tool descriptions, and retrieved evidence. A context window is the bounded capacity available for that material, measured in tokens, the units models process. Anthropic describes context engineering as selecting and maintaining what enters this working input. Anthropic’s context-engineering explanation.
Think of context as the working packet for the next decision. A file sitting elsewhere in the application may be accessible through a tool, but its contents are not necessarily in that packet.
Memory: what the application preserves
Memory terminology varies by system. LangGraph, for example, distinguishes conversation-scoped state, which can be persisted so a thread can resume, from long-term memory shared across sessions. “Short-term” therefore need not mean “deleted when the program stops.” LangChain memory overview.
For a product evaluation, ask separately what is stored, how long it remains, which conversations can use it, and how it is selected for recall. A persistence feature alone does not answer those questions.
Retrieval: how information is found again
Retrieval is the lookup step. It can bring a saved memory or an external source into the current task. In retrieval-augmented generation, usually shortened to RAG, retrieved material helps inform the generated response. The original RAG paper studied a particular trained architecture; it does not validate every product now using that label. Lewis et al..
The categories overlap in a useful way: memory describes what survives; retrieval describes how it is accessed; context describes what is available when the model responds.
Decide by how the information changes
AMF recommends assigning information a treatment based on its role, rather than putting everything into one growing history.
- Current instructions and constraints: Include them in the working context. Make exceptions and unresolved questions explicit.
- Useful continuity: Preserve a concise record of decisions, completed steps, and preferences where appropriate.
- Changing facts: Keep a source reference and retrieve the current value when the decision depends on it.
- Unsupported interpretations: Label them as provisional. Do not quietly promote a generated guess into an established fact.
This is an AMF design recommendation, not a universal storage standard. The distinction depends on the task. An old policy can be the correct evidence for explaining a historical decision, while being unsuitable for deciding a new request.
Before storing something, ask whether it will help future work and whether retaining it is appropriate. Before reusing it, ask whether its age or scope could change the answer.
A hypothetical equipment request
Imagine an internal assistant preparing an equipment recommendation for a new employee. Its assignment is to draft a recommendation, not place an order.
The following facts are invented for this example:
- The current request specifies a portable laptop for frequent travel.
- A saved team note says the standard choice was Model A under catalog version 4.
- A lookup of the current catalog, version 5, shows Model A has been withdrawn and Model B is now standard.
- A current inventory lookup shows Model B is unavailable.
- The assistant has not found the policy governing substitutes.
A useful decision record would look like this:
| Information | Treatment | Effect on this recommendation |
|---|---|---|
| Frequent travel is part of the current request | Keep in working context | Preserve portability as a requirement |
| Model A was previously standard | Retain as dated history, not current guidance | Explains the old note without controlling the new answer |
| Catalog version 5 names Model B | Use retrieved policy evidence with its version | Identifies the current standard |
| Model B is unavailable in the inventory lookup | Record the observation and lookup time | Blocks an immediately available standard recommendation |
| Substitute eligibility is unknown | Keep the gap explicit and look up the relevant rule | Do not invent an approved alternative |
The resulting recommendation is conditional: “Model B is the current standard, but the inventory check shows it unavailable. Substitute eligibility remains unresolved.”
If the task resumes tomorrow, the assistant should recover the travel requirement, the recommendation’s status, and the missing policy question. AMF would also require a fresh stock check before relying on availability. Saving “Model B unavailable” as a timeless fact would carry yesterday’s observation into tomorrow’s decision.
This example demonstrates information handling, not a tested implementation or evidence that any particular agent would behave correctly.
Why a bigger context window is only part of the answer
A larger working packet can hold more material. It does not establish that the model will use every relevant detail correctly.
Liu and colleagues’ Lost in the Middle study tested multi-document question answering and key-value retrieval. Performance depended on where relevant information appeared, with weaker results often occurring when it was in the middle. These results concern the study’s models and tasks; they do not establish the performance of every current model. Lost in the Middle.
Applications may also shorten a long conversation through compaction: replacing detail with a summary. Anthropic describes this technique and warns that aggressive compaction can lose information that later proves important. Context management for longer tasks.
AMF’s recommendation is to protect the decision’s conditions during that shortening. In the equipment example, “Recommend Model B” is an inadequate summary. It drops both the stock problem and the unresolved substitute rule.
Choose the simplest arrangement that fits
A one-off question over a supplied document may need only the document and clear instructions. A recurring assistant may benefit from saved preferences and task state. A decision involving changing operational facts needs a way to consult the relevant source again.
These options can coexist. Do not assume that every workflow needs a separate memory database or elaborate retrieval system.
The sources here establish useful concepts and specific implementation or research findings. They do not determine the best retention period, storage design, or retrieval settings for your organization. Those choices need to reflect access boundaries, the consequences of stale information, and the work people must check.
Next action
Take one recurring task and list its decisive inputs. For each, record:
- Where the information comes from.
- What must be visible at decision time.
- What should persist for continuity.
- What event or elapsed time requires another lookup.
- What the agent should leave unresolved if the source is missing.
Start with a record a colleague can read. If the team cannot explain those five points, adding more memory is unlikely to clarify the decision.
Provenance
Prepared for Agent Model Fit with AI-assisted source inspection, synthesis, and drafting on November 6, 2026. Research claims are attributed to the cited papers; implementation descriptions are attributed to the cited vendor documentation. The information-treatment guidance and equipment decision record are AMF synthesis. The scenario is hypothetical, and no implementation trial, measured business outcome, or human review is claimed.