Skip to main content

AI Agent Runtime: Understanding Agent Memory – Context, Working Memory, and Conversation History

Calculating read time…

In an order-support workflow, a customer might ask, "What is the status of my order?" and then follow up with, "What about the delivery date?" The second question only makes sense when the agent can relate it to the first. But there are two separate engineering concerns: what information is available to the model while it is processing a request, and what information can be carried into a later agent run.

Agent memory is the runtime's ability to retain and reuse information. The key distinction is between context, the information available to the model for its current request, and stored conversation history, saved interactions that the runtime can draw on later. Working memory describes the execution state used while completing a task; persistence describes saving state beyond the current in-memory processing step. These ideas are related, but they are not interchangeable. Oracle Integration 3 (OIC Gen3) provides a concrete example: a Conversation ID associates separate agent runs so earlier interactions can support a follow-up. That continuity does not guarantee indefinite retention, make old business facts current, or train the model.

1. What You Will Build

This article explains agent memory through an illustrative Order Support Agent. You will distinguish the information available to the current model request, saved conversational interactions, and authoritative business records:

  • Current model request: Instructions and relevant information supplied to the model during its current processing step.
  • Conversation history: Information associated with previous agent runs through a shared OIC conversation ID.
  • Business systems: Authoritative order, shipment, payment, and delivery records.

The practical exercise uses two customer messages about one synthetic order. You will examine how a conversation ID links separate agent runs, understand what Oracle documents about retained interactions, and decide which facts must be retrieved again from a business system.

Completion evidence: A classification table, a two-run conversation trace, and verification tests that distinguish conversational continuity from verified business state.

2. In This Post

3. Before You Begin — Shared Setup

3.1 Documentation baseline

Verify your actual OIC instance version, maintenance level, and available action operations before implementation. A release-note minimum is not proof that a capability is enabled in a particular tenant.

Sources: Using Agentic AI in Oracle Integration 3, OIC 26.10 release notes, OIC 26.01 release notes, OIC 26.01.2 release notes, Oracle Integration 26.04 product update, the 26.04 product working document, and the OIC 26.07 release notes.

3.2 The example business scenario

Assume an enterprise has an Order Management System (OMS) containing customer orders and shipment information. OIC exposes a read-only business operation that retrieves current order status. An OIC project AI Agent uses that operation when answering customer questions.

We will use this synthetic order:

Worked example — Order ORD-10482

Business fieldIllustrative value
Order IDORD-10482
Customer referenceCUST-204
Initial statusSHIPPED
Initial estimated delivery12 November 2026
Tracking referenceTRK-7781

All identifiers and values are fictional test data. They are not connected to an actual customer or order.

The example assumes that the order-status operation already exists, or will be designed separately, as a read-only tool available to the AI Agent. Its input and output fields below are illustrative application contracts, not automatically generated OIC AI Agent mapper schemas. This article does not define the OMS API, authentication, agent-tool registration steps, or a deployed connection.

3.3 The important identifiers

Three identifiers must not be confused:

Identifier Purpose
Order ID Identifies the actual business order in the OMS.
Conversation ID Associates multiple agent runs with one conversation.
Agent instance ID Identifies an individual runtime execution.

Oracle explicitly documents that one conversation ID can span multiple instance IDs. Therefore, a conversation and an individual agent execution are not the same unit.

Reference: Invoke an AI Agent from an Integration.

4. Understanding Context, Working Memory, and Stored History

4.1 Current model context

Model context is the collection of instructions and information supplied to a Large Language Model (LLM) for a particular model request. It is the information the model can use while producing that response.

In a general agent design, the context may contain configured instructions, the current user message, relevant previous conversational messages, and results returned by tools. This is a conceptual description, not a published specification of OIC's exact prompt assembly.Context is not itself a permanent database.

Kid-friendly analogy

Imagine a student taking an examination. The student has a question paper and a limited amount of space on the desk. Only the books and notes placed on that desk can help answer the current question.

In this OCI AI integration, this means: The model can use the instructions and information supplied in its current context. Information stored somewhere else must be retrieved or supplied before the model can use it.

There is an important distinction between a conversation existing in OIC and its contents being available to the model at a particular moment. Stored information is useful only when the runtime makes relevant information available to the current model request.

The model also has a finite context window. A larger history does not mean unlimited information can be processed in one request. Context capacity depends on the selected model and the actual request composition.

4.2 Working memory during execution

Working memory is a useful general AI Agent design term for the temporary information available while an agent is carrying out its current task. It may include the latest user request, the action selected by the agent, intermediate tool results, and information needed to decide what to do next.

Consider the order-status example:

  1. The customer asks for the status of ORD-10482.
  2. The agent selects the order-status tool.
  3. The tool returns the current status and estimated delivery date.
  4. The agent uses that observation to prepare its answer.

These intermediate results belong to the execution process. They help the agent complete its current task, but the phrase "working memory" does not automatically imply a separate durable storage resource. Working memory and model context overlap: some execution state is supplied to the model, while other state is used only by the orchestrator, the software coordinating model calls and tool calls. Saving that state is a separate operation.

Kid-friendly analogy

Imagine a child solving a puzzle. They remember which pieces they have already tried and which piece fits the current gap. That temporary thinking helps complete the puzzle, but it does not mean the child has saved every step in a permanent notebook.

In this OCI AI integration, this means: Tool observations and intermediate decisions can help the agent finish its current execution. Do not treat that temporary execution state as a guarantee of cross-run persistence.

OIC-specific boundary: Oracle's Test Agent documentation exposes activity such as System, User, Tool, and Assistant messages. This gives developers an observable view of the agent execution. However, Oracle does not describe the complete internal representation of its working-memory mechanism in the conversation-history documentation. We should not invent internal runtime data structures or assume a separately configurable working-memory database.

Reference: Run and Test the AI Agent.

4.3 Stored conversation history

Conversation history is information associated with earlier interactions that can be made available during a later interaction.

This is where the OIC Conversation ID becomes important. Oracle documents that an AI Agent receives a conversation ID automatically when it runs, unless you supply one. You can specify the same ID in multiple agent runs to associate those requests with one conversation.

Oracle describes a conversation as including interactions such as prompts, data, and decisions. It also states that the agent can use the conversation history of other runs that share the same ID. 

Kid-friendly analogy

Imagine a librarian giving you a notebook with a unique reference number. When you return another day and present the same number, the librarian can connect your new visit with the earlier visit recorded in that notebook.

In this OCI AI integration, this means: Use the same Conversation ID across agent runs that belong to the same customer conversation so the agent can access previous conversational information.

Here, persistent means that OIC can associate a later run with earlier runs when the same Conversation ID is supplied. It does not mean one integration invocation stays active, nor does it mean that the history is retained forever. Oracle points to the OIC service-instance data-retention setting for AI Agent data; the exact record-level retention behavior is discussed later.

For example:

  • Run 1 creates the conversation.
  • Run 2 starts with the same conversation ID and can use the previous conversational history.
  • Run 3 can continue the same conversation if the same ID is supplied.

These are separate executions associated with one conversation. The conversation ID is therefore a link between runs, not proof that one integration instance has remained continuously active.

Important: A conversation ID is not the same as a customer identity, authentication credential, or business authorization token. Your application must control who is allowed to reuse a particular conversation ID. Do not share one conversation ID across unrelated customer sessions.

4.4 ReAct and persistent-memory activities

ReAct is an agent pattern that combines model-guided decisions with actions, such as calling a tool, and observations, such as the tool's response. It may repeat this cycle before answering. It does not require you to see the model's private reasoning. Oracle's OIC tutorial shows a system-generated ReAct pattern integration as part of an agent's execution.

To understand memory in this cycle, distinguish three jobs:

Runtime jobWhat it means
Maintain working stateKeep messages and relevant tool observations available while the agent completes its current task.
Read or write saved stateRetrieve previously saved information or persist updated information beyond the current in-memory processing step.
Prepare model contextSupply instructions, the latest request, and relevant available information to a particular model call.

These are conceptual responsibilities, not OIC menu labels or a guaranteed sequence for every generated pattern. When you encounter an activity described as persistent memory in a ReAct integration, the relevant question is whether it reads saved interaction state, writes updated state, or both. Its name alone does not establish which messages are saved, when the write succeeds, what is restored later, or how much information reaches the next model call.

Kid-friendly analogy

A student works with notes on a desk and saves useful notes in a notebook before leaving. At the next lesson, the student opens the notebook and places the relevant notes on the desk. The saved notebook and the notes currently on the desk serve different purposes.

In this OCI AI integration, this means: Saving interaction state can support later reuse; supplying that information to a model call makes it part of current context. Saving everything does not mean the model can read everything at once.

For the order example: A later run can use earlier interaction information to identify ORD-10482. Its current model context must contain enough relevant information to interpret the follow-up. A fresh order-status tool result is still needed to establish the latest delivery estimate. Persisting the earlier answer does not refresh that estimate.

Use this when: You need to distinguish keeping information during a task, saving it for later, and making it available to the model now.

References: Run and Test the AI Agent and Run an AI Agent with Knowledge of Previous Runs.

5. How OIC Project AI Agent Conversation History Works

5.1 What happens when an agent runs?

Oracle's Test Agent documentation provides a useful way to understand the observable execution. When you inspect a test run, the activity view displays several message categories.

Activity category What it represents
System The system instructions sent to the LLM. Oracle's example identifies the agent role, agent guidelines, and agent pattern guidelines.
User The user message submitted to the agent.
Assistant Oracle describes the Assistant activity as showing the agent's reasoning. It can include the assistant's response or decision in the displayed run activity; do not interpret this view as a complete internal model trace or a guaranteed record of every internal operation.
Tool The response returned by an invoked tool.

This distinction helps identify the different types of information involved in an agent conversation.

For example, the agent can receive the instruction to use an order-status tool, receive the customer's question, invoke the tool, and use the returned result to formulate a response.

These activity categories are observable representations. They should not be interpreted as a published database schema or a guarantee that every internal model operation, complete tool payload, or private reasoning trace is retained and replayed unchanged.

5.2 What information can be retained?

The following table separates documented conversation content from the details that still require caution.

Information Interpretation
Agent prompts and user messages Prompts and data are explicitly identified as conversation interactions.
Assistant responses and decisions Decisions are explicitly included in the conversation description; assistant activity is also visible in Test Agent.
Tool responses Tool output is visible in the documented execution activity. The conversation documentation does not provide a field-by-field persistence contract for every tool result.
Tool request arguments The Test Agent activity documentation describes tool activity, but the conversation-history documentation does not define the exact persisted representation of each tool request argument or guarantee that the same arguments are replayed into every later model request.
System instructions The current system message is visible in the test activity. Historical version retention and replay of system instructions are not separately specified.
Business database records These belong to their respective business systems, not to conversation history as an authoritative record.

There is a difference between information visible in a runtime trace and a documented guarantee about what the history store retains or places into every subsequent model request. That distinction matters when designing auditing and recovery.

5.3 How to inspect the conversation ID

You can inspect the conversation ID in the Test Agent experience:

  1. Open the required OIC project.
  2. Select Design.
  3. Select AI Agents in the left navigation.
  4. Locate the AI Agent and select Actions → Run.
  5. Run the agent and select the test run.
  6. Inspect the Activity section for the Conversation ID.

Oracle's conversation-history procedure also points to the project's Observe tab and the Agent instances page. For a controlled test, use the value explicitly labelled Conversation ID in Test Agent Activity and compare the values shown in your tenant before relying on the Observe-page column.

To continue the same conversation in the Test Agent page:

  1. Open the agent's Test Agent page again.
  2. Click the + icon to open the Conversation ID dialog.
  3. Supply the previously obtained conversation ID.
  4. Click Submit.
  5. Enter the next prompt and click Run.

The agent can use the conversation history associated with other runs that use the same ID. The exact details made available to a later model request are not a field-by-field replay contract; verify the behavior in Test Agent Activity for the relevant scenario.

Concurrency restriction: Oracle documents that starting another run with a conversation ID that is already in progress produces an error. The same restriction applies when the conversation has the status Waiting for tool to process. Design your caller to serialize turns for one conversation or otherwise coordinate their execution.

Official procedure: Run an AI Agent with Knowledge of Previous Runs.

6. Practical Prototype: Two Runs, One Conversation

Now apply these concepts to the Order Support Agent. The purpose of the prototype is to demonstrate how conversational context and business information work together without treating them as interchangeable.

6.1 Define the application inputs

Assume an OIC integration already has an inbound trigger whose request schema contains the application fields shown below. The trigger type, REST resource path, authentication, and schema-generation steps are outside this memory-focused example; they must be designed and verified for the actual integration rather than inferred from the AI Agent action.

For this exercise, define the following application-owned inbound contract. These are example integration fields, not native OIC AI Agent action properties.

{
  "conversationKey": "SESSION-ORD-10482-01",
  "message": "What is the status of order ORD-10482?"
}

Here:

  • conversationKey is an opaque application-generated reference for one customer conversation.
  • message contains the latest user question.

Use an opaque identifier rather than embedding a customer's name, email address, or other sensitive information into the conversation ID. The sample uses an application-owned key, but the OIC procedure cited here does not publish a complete accepted-character or maximum-length contract for caller-supplied IDs. Treat the sample value as illustrative and validate the actual accepted format in your instance. Protect the mapping between the authenticated customer session and this value in the application; possession of an ID must not be treated as authorization.

6.2 Configure the native OIC AI Agent action

Oracle documents a native AI Agent action that allows an integration to start an active AI Agent belonging to the current or a shared project.

For this prototype, use an OIC integration with an inbound trigger that receives the application-owned JSON contract above.

Follow the documented action configuration:

  1. Open the integration canvas.
  2. From the Actions panel, drag an action onto the canvas, or click the add icon and search for AI Agent.
  3. In Configure Basic Info, enter the required endpoint name and, optionally, a description.
  4. In Configuration, select the current project or a shared project that contains the agent. Select the required AI Agent. Only agents in the Active state are available for selection.
  5. Click Continue, then click Finish to complete the action configuration. OIC automatically creates a map.
  6. Open the generated mapper. Map the application conversation key to the Conversation Id parameter and the latest user message to the Prompt body field.

The consequential mapping is:

Inbound application field Native OIC AI Agent target Purpose
conversationKey Parameters → Conversation Id Associates this execution with the intended conversation.
message Body → Prompt Supplies the latest natural-language request.

Project Id and Agent Id are optional parameters when the project and agent have already been selected during configuration. The documentation specifies that you can provide either a Prompt or a Prompt Template Id with Payload. This example uses the direct Prompt route.

Click Validate in the mapper, return to the integration canvas, click Save, and then activate the integration. These are the documented completion steps for the native action.

Important distinction: The application-owned field conversationKey is mapped to the actual OIC parameter Conversation Id. They are not interchangeable product field names.

Official procedure: Invoke an AI Agent from an Integration.

6.3 First execution — customer asks for order status

The first incoming request is:

{
  "conversationKey": "SESSION-ORD-10482-01",
  "message": "What is the status of order ORD-10482?"
}

The integration passes the message to the AI Agent and maps the session's application-owned conversation key to the native Conversation Id parameter. The sample identifier is illustrative; the cited OIC procedure explains that a caller may supply an ID but does not specify a complete format or length contract. Verify any tenant-specific validation before adopting an identifier-generation scheme. If you omit the parameter, OIC generates an ID, but the application must then obtain and preserve that generated value through a supported response or correlation path before it can deliberately reuse it.

For the worked example, the agent uses a read-only tool with this application-defined contract:

Tool: get_order_status

Input:
{
  "orderId": "ORD-10482"
}

Illustrative result:
{
  "orderId": "ORD-10482",
  "status": "SHIPPED",
  "trackingNumber": "TRK-7781",
  "expectedDeliveryDate": "2026-11-12",
  "source": "ORDER_MANAGEMENT_SYSTEM"
}

The example assumes the tool is registered and configured as an agentic AI tool. Its name, input schema, response schema, and business behavior are tutorial-defined rather than native OIC conversation-history properties.

The agent can now formulate an answer such as:

Illustrative assistant response

Order ORD-10482 has been shipped. The current estimated delivery date provided by the Order Management System is 12 November 2026. You can use tracking reference TRK-7781 to check the shipment.

This response is based on the illustrative business result, not on an assumed model output or an actual execution.

After this run, the conversation has information connecting the customer's question with the order and the response made by the agent.

6.4 Second execution — customer asks a follow-up

Now the customer sends a second request:

{
  "conversationKey": "SESSION-ORD-10482-01",
  "message": "Can it arrive before Friday?"
}

Notice that the latest message does not repeat the order number. It uses the word "it". The previous conversation supplies the reference needed to understand the follow-up.

For the date comparison below, assume the customer asks this question on Monday, 9 November 2026. In this example, "Friday" therefore means 13 November 2026.

Because the application again maps the same conversation key to Conversation Id, OIC associates this new run with the earlier conversation. However, the agent must not blindly reuse the earlier delivery date. We will assume the OMS has subsequently received a carrier update:

{
  "orderId": "ORD-10482",
  "status": "SHIPPED",
  "trackingNumber": "TRK-7781",
  "expectedDeliveryDate": "2026-11-14",
  "source": "ORDER_MANAGEMENT_SYSTEM"
}

The agent should use the conversation history to understand which order the customer means, but the current order-status result to answer the delivery question.

Illustrative assistant response

The latest estimated delivery date for ORD-10482 is 14 November 2026. That is after Friday, 13 November, so the current OMS estimate does not indicate delivery before Friday.

The important engineering result is that the agent has used two different sources of information correctly:

  • Conversation history resolved the reference "it".
  • The OMS supplied the current delivery estimate.

Business rule for this prototype: The agent instructions should require a fresh read-only order-status lookup before making a claim about current order status or delivery estimates. This is an application-defined rule. Do not assume that supplying a Conversation ID automatically refreshes business data or enforces this rule.

6.5 What should be inspected?

When you run the prototype, use the Test Agent activity to inspect the visible execution sequence:

  • System: Check the current agent role and guidelines.
  • User: Check the actual question submitted for that run.
  • Tool: Check the returned order result and whether the tool was invoked.
  • Assistant: Check whether the final answer reflects the relevant order result.

Also inspect the corresponding OMS record. A response that repeats the previous expected delivery date without obtaining the latest result is not evidence of correct business processing.

Oracle documents a separate Get AI Agent Activity Stream operation in the native AI Agent action. It accepts the Conversation ID and returns an activity stream; the response includes information indicating the agent's current status. It is an observation mechanism, not a substitute for checking the business system or a guarantee that all historical model inputs are exposed.

To configure it, add an AI Agent action to an integration. In Configure Basic Info, enter the required endpoint name and, optionally, a description. In the Action field, select Get AI agent activity stream, then click Continue. On the Summary page, click Finish. In the generated mapper, map the Conversation ID, then return to the canvas, save, and activate the integration.

Reference: Get the Activity Stream for an AI Agent.

7. Conversation History vs Knowledge Base vs Model Training

One of the most important design decisions is deciding which capability actually solves the information problem.

Conversation history, Retrieval-Augmented Generation (RAG), business databases, and model training can all make information available to an AI application, but they do so in different ways.

Kid-friendly analogy

Imagine a school. A student's notebook records previous conversations with a teacher. The library contains reference books. The school register contains official records. Training teaches the student new skills rather than simply adding another page to the notebook.

In this OCI AI integration, this means: Conversation history provides continuity, a knowledge base supplies searchable document evidence, business systems hold authoritative records, and model training changes the model rather than preserving one customer's chat.

7.1 Conversation history is not a knowledge base

A knowledge base is a collection of documents or other information made searchable for an AI workflow. In a RAG architecture, relevant source material is retrieved to support the current answer.

OIC provides a project-level Knowledge Base capability. Before adding one, Oracle's procedure requires a configured connection to OCI Search with OpenSearch. During ingestion, documents are split into chunks and the selected embedding model computes vector embeddings for those chunks before they are uploaded to the connected knowledge base. Oracle documents the KnowledgeBaseSearch integration for querying the knowledge base; Object Storage ingestion can use the ObjectStoreKnowledgeBaseIngestion integration. The embedding model, connection, and ingestion choices are separate from Conversation ID-based conversation history.

For example, a shipping policy document could contain:

Shipping Policy: Standard ground shipments have an estimated transit window of two to five business days. An estimated delivery date is not a guaranteed delivery commitment.

The agent could retrieve this policy when explaining what an estimated delivery date means.

That is different from remembering the earlier exchange about ORD-10482.

Question Appropriate source
Which order were we discussing? Conversation history
What is the latest delivery estimate? Order Management System
What does the shipping policy say? Knowledge base or authoritative policy source

A stored conversation is not automatically converted into embeddings or indexed for semantic search just because it is persistent.

References: FAQs for Knowledge Base and Add a Knowledge Base.

7.2 Conversation history does not automatically extract business facts

Suppose a customer says, "Please notify me by SMS next time."

The conversation may contain that statement. However, that does not prove that the preference has been extracted, validated, converted into a structured customer preference, or saved in the CRM.

Those operations require an explicit application design.

For example:

  1. Identify the customer's statement.
  2. Determine whether it represents a persistent preference or only a one-time request.
  3. Obtain confirmation when required by the business process.
  4. Update the authorised customer-preference record.
  5. Verify the saved value in the CRM.

The historical conversation can provide context for this workflow, but it does not perform structured fact extraction and persistence automatically.

7.3 Conversation history does not train the model

Model training changes the learned parameters of a model through a training or fine-tuning process. Conversation history, by contrast, provides previously recorded interaction information as context for a later response.

Consider the difference:

  • When the customer asks a follow-up question with the same Conversation ID, the agent can use prior conversational information.
  • That does not mean the underlying LLM has permanently learned the customer's order history.
  • It does not mean the model's weights have been updated using the conversation.

Do not describe a successful conversational follow-up as model learning. The documented OIC feature is conversation continuity, not automatic fine-tuning or model training.

8. Retention, Stale Information, and Context Growth

8.1 What does persistent actually mean in OIC?

Oracle explicitly states that AI Agent data is retained according to the data retention period of the Oracle Integration service instance.

The instance administration documentation publishes these default retention values:

Instance edition Documented default Changeable?
Standard 32 days No
Enterprise 32 days Yes
Healthcare 184 days No

The OIC AI Agent guide explicitly says AI Agent data is retained according to the Oracle Integration service-instance data-retention period. The administration guide states that, when an Enterprise instance's retention period is decreased, data older than the newly selected period is deleted when the selection is saved.

There is an important scope qualification: the administration page says the numeric retention setting described there applies to production tracing data; it lists separate retention periods of 24 hours for Debug tracing and 8 days for Audit tracing. The AI Agent guide points to the instance setting, but the public pages do not define a separate conversation-record schema or explicitly map every conversation message, tool argument, or internal record to a specific trace category and deletion schedule. Treat the table as the published service-instance setting referenced by the AI Agent guide—not as a per-message retention SLA. Confirm the exact treatment of conversation records with Oracle before using this feature as a regulated retention control.

To inspect the setting:

  1. Open the Oracle Cloud Console.
  2. Navigate to Developer Services → Application Integration → Integration.
  3. Open the relevant Oracle Integration instance.
  4. On the instance details page, locate Data retention.
  5. Click Manage to inspect the available retention selection.

Reference: Edit the Data Retention Period.

8.2 More history is not always better

A conversation can become longer as the customer keeps asking questions. More historical information may help with continuity, but it can also introduce unnecessary content into the model's context.

For example, a long conversation might contain:

  • Old delivery estimates.
  • Orders that have already been completed.
  • Repeated questions and answers.
  • Temporary customer instructions.
  • Personal information that is no longer necessary.
  • Tool results that have become obsolete.

If all that information is treated as equally relevant, the agent may have difficulty identifying which facts matter to the latest request.

8.3 Prevent stale information from becoming business truth

Conversation history can preserve a response that was correct when it was produced but is no longer correct.

For example, the agent may have told the customer that delivery was expected on 12 November. A later carrier update could change the estimate to 14 November.

When the customer asks another delivery-related question, the agent should obtain the latest delivery information from the OMS.

Use this simple rule:

Use conversation history to understand what the customer means. Use the authoritative business system to establish what is true now.

8.4 Avoid unnecessary sensitive information

A conversational history may contain customer messages and associated data. Therefore, design the application to minimize what enters the conversation in the first place.

For example, an order reference may be sufficient to identify an order. The agent normally does not need an entire payment-card number, password, authentication token, or unrelated personal information to answer a shipment question.

Use appropriate controls:

  • Pass only the business fields required by the current task.
  • Avoid putting personal information into conversation identifiers.
  • Restrict reuse of a conversation ID to its intended customer or session.
  • Avoid copying complete sensitive tool responses into assistant messages.
  • Review logging, monitoring, and data-retention requirements.
  • Keep authoritative customer and order records in their governed business systems.

A conversation ID is a correlation mechanism, not an access-control mechanism. Your application must enforce the relationship between the authenticated caller, customer, session, and permitted data.

8.5 What durability guarantees are not established?

The public documentation establishes the use of a conversation ID, the ability to associate multiple runs, and a link to instance data retention. That is useful documentation, but it is not a complete storage or disaster-recovery specification.

The public documentation cited here does not establish:

  • The internal database schema or physical storage implementation used for conversation history.
  • Cross-region replication or backup guarantees specifically for conversation records.
  • Point-in-time recovery guarantees for individual conversation messages.
  • The exact deletion timing of every individual conversational or tool-related record.
  • Whether every full tool request and observation is replayed into each later model request.
  • A guaranteed automatic history summarization or truncation policy for OIC project AI Agents.
  • Retention beyond the documented instance retention period.

These are meaningful production questions. A conversation ID should not be used as the sole durable system of record for an order, approval, payment, or long-running business process.

9. Verification and Troubleshooting

The purpose of testing is not merely to prove that the agent can generate a second answer. It is to verify that the correct information comes from the correct source.

Test case Expected behaviour Evidence to inspect
Same Conversation ID across two sequential runs The agent can relate the follow-up to the previous order discussion. Same Conversation ID and relevant previous interaction in the Test Agent activity.
New Conversation ID The new customer session should not be deliberately associated with the previous session's history. Different conversation IDs and isolated session mapping.
Conversation ID omitted OIC generates a conversation ID automatically; do not expect the caller to continue the earlier conversation. Conversation ID displayed for the new run.
OMS estimate changed between turns The latest answer uses the updated OMS value rather than an old conversational response. Tool invocation, tool result, and current OMS record.
Same conversation invoked while in progress OIC reports the documented conversation-in-progress error. Agent status and invocation error.
Same conversation in Waiting for tool to process Another run cannot be started for the same conversation while it is in this state. Agent activity status and invocation error.
Unrelated users attempt to reuse one ID The application should reject unauthorized cross-session reuse. Authenticated user-to-session mapping and authorization decision.

Execution status: Not executed. These are prescribed tests and expected outcomes, not results from a live OIC tenant.

9.1 Common troubleshooting scenarios

Symptom Likely cause and correction
Agent does not understand a follow-up question. Check whether the second run received the same Conversation Id. Also inspect whether the previous interaction is available in the test activity.
A different order's history appears. Check the application session-to-conversation mapping. Do not reuse a conversation ID across unrelated customers.
The agent repeats an old delivery estimate. Inspect the tool activity. Make sure the current order status is retrieved and the new result is used in the answer.
The next run fails when using the previous ID. Check whether the earlier conversation is still In Progress or Waiting for tool to process. Coordinate requests instead of launching overlapping runs.
History cannot be continued with an older agent through its API endpoint. For agents created before 26.04 and invoked through an API or AI Agent endpoint, check the documented Conversation ID known issue and its URL query-parameter workaround. This note does not establish a workaround for the native integration action or Test Agent.
A 404 occurs, but an agent instance record still appears. Oracle documents that an incorrect AI Agent endpoint URL can return 404 while still creating a conversation record and AI Agent instance record, leaving the agent instance In Progress. Validate the endpoint path before invoking it; do not treat the record alone as evidence of a successful run.

Reference: Agentic AI Issues.

10. Completion Checklist and Next Steps

Use the following checklist to evaluate whether the memory architecture is correctly understood and designed.

Architecture checklist

  • Current model context is distinguished from stored conversation history.
  • Working memory is understood as execution-time state, not automatically durable storage.
  • The same Conversation ID is used for the intended follow-up interactions.
  • Individual agent instance IDs are not confused with the business order ID.
  • Customer sessions have an application-enforced isolation policy.
  • Live order facts come from the authoritative OMS.
  • Knowledge base retrieval is used for document evidence, not as a substitute for live order status.
  • Conversation history is not treated as automatic fact extraction or model training.
  • Retention is checked against the actual OIC service-instance settings.
  • Concurrent runs and waiting-tool states are handled.
  • Historical sensitive information is minimized and governed.
  • Tests inspect both conversational continuity and business-data freshness.

What remains outside this prototype?

The example explains the distinction between current context, saved interaction state, and live business records. It does not establish the internal storage implementation, guaranteed history compaction, or a governed store of long-term business facts.

Those requirements would need separate design work. In particular, when an application needs structured preferences, long-lived customer facts, or recoverable workflow checkpoints, it should persist them in an appropriate governed data layer and retrieve them through authorised application operations.

11. Five Useful FAQs

1. Does an OIC AI Agent remember the previous question automatically?

OIC generates a Conversation ID automatically when an agent runs. To deliberately associate separate runs and provide continuity, pass the same Conversation ID into the subsequent run. A newly generated ID does not establish the association with an earlier conversation.

2. Is Conversation ID the same as Agent instance ID?

No. An agent instance identifies an individual execution, whereas a conversation can span multiple instance IDs. Use the Conversation ID for conversational correlation and the relevant business identifier for the actual business record.

3. Does conversation persistence mean the model permanently learns everything?

No. Conversation history supplies previous interaction information. It does not automatically modify the model's trained parameters, extract permanent business facts, or create new model knowledge through training.

4. Can an agent answer an old order question using conversation history without calling the business system?

It may be able to understand which order the customer means using historical context, but the application should obtain current transactional information from the authoritative business system. Conversation history can contain information that is stale or no longer correct.

5. Does OIC automatically summarize or truncate long conversation history?

The OIC documentation examined for this tutorial does not establish a specific project AI Agent setting for automatic history summarization or configurable token-budget truncation. Do not assume these controls from a different Oracle agent product apply to OIC project AI Agents.

12. Additional References

The following Oracle sources were inspected for the OIC-specific procedures, release notes, feature boundaries, retention information, and product distinctions used in this article. A source publication or minimum-version entry does not prove that a feature is enabled in a particular tenant.

  1. Using Agentic AI in Oracle Integration 3 — Current guide G41284-23, October 2026.
  2. OIC 26.10 Release Notes — September 2026 release baseline used for this review.
  3. OIC 26.01 Release Notes — Conversation ID in AI agents, minimum version 26.01.
  4. OIC 26.01.2 Release Notes — Invoke an AI agent from an integration, minimum version 26.01.2.
  5. New Adapters and Connectivity Enhancements in Oracle Integration 26.04 — Product Manager post describing the native AI Agent Invoke action as introduced in 26.04; its date differs from the formal 26.01.2 release-note entry.
  6. OIC 26.07 Release Notes — Lists Get AI Agent activity stream as a new AI Agent action operation, minimum version 26.07.
  7. What's New in Oracle Integration 26.04 — Product Manager working document that says the native action supports Get Agent Activity Stream in 26.04; the post carries a working-document disclaimer.
  8. What's New in Oracle Integration 26.07 — States that 26.07 adds agent run status to the Get Agent Activity Stream response.
  9. Run an AI Agent with Knowledge of Previous Runs — Conversation ID, history association, how to obtain/reuse the ID, and concurrency restrictions.
  10. Invoke an AI Agent from an Integration — Native action prerequisite, configuration, parameters, and Conversation Id/Prompt mapping.
  11. Run and Test the AI Agent — System, User, Assistant, and Tool activity descriptions.
  12. Get the Activity Stream for an AI Agent — Activity-stream action configuration and Conversation ID input.
  13. Edit the Data Retention Period — Published service-instance retention settings and their tracing-data scope.
  14. Agentic AI Issues — Conversation ID compatibility issue for agents created before 26.04 and the 404/instance-record issue.
  15. FAQs for Knowledge Base — OIC Knowledge Base ingestion, chunking, embeddings, and search integrations.
  16. Add a Knowledge Base — Project-level Knowledge Base setup and OCI Search with OpenSearch connection prerequisite.
  17. Agent Memory — Oracle AI Data Platform — A separate Oracle product with its own context, history-limit, truncation, stored-memory, and session-retention controls. These are not OIC Project AI Agent settings.
  18. Agent Memory Is Not RAG: Conversation IDs, Durable State, and Scoped Recall — Oracle Developers architectural discussion. Used for general design principles, not as proof of OIC runtime behavior.

13. Finally — Summary

Context is what the model can use now; stored conversation history is saved interaction information that the runtime can reuse later. Working memory supports the current task, and persistence saves state beyond the current in-memory processing step. In a ReAct pattern, maintaining messages, saving interaction state, and supplying context to a model call are related responsibilities with different purposes. OIC documents Conversation ID-based history for associating separate agent runs. A persistent-memory activity does not, by its name alone, establish how storage works or guarantee unlimited recall.

However, conversational continuity does not provide semantic search, structured fact extraction, automatic summarization, or model training. It also does not make an earlier business result permanently authoritative. The OIC instance retention setting is the documented reference for AI Agent data, but the administration page's published numeric periods are described for production tracing data; do not assume they are a per-message conversation-history deletion schedule.

For enterprise integrations, use conversation history to understand the customer, a knowledge base to retrieve document evidence, and business systems to establish current transactional facts. The most important design principle is simple: remember the conversation, but verify the business state.

Comments