Manage Agent Caches

Beginning in Strategy One (January 2026), user questions and answers are cached for each agent, and you can manage these caches.

Agentic caching improves both the consistency and performance of answers generated by agents. It operates at the individual agent level, allowing fine‑grain control over what is cached and how cached content is reused.

With agentic caching:

Responses can be partially cached, meaning individual components of a multi‑step answer can be reused independently.
Cache entries are associated with a specific agent, not a specific user, enabling cache reuse across all users interacting with that agent.

The caching mechanism relies on semantic similarity detection. When users ask questions, the system evaluates how similar each new question is to existing ones and groups them into cache groups, also called buckets. All questions in a bucket are considered variations of the same intent and can share the same or closely related underlying logic (for example, the same SQL).

Because caching is done at the agent level, a cache entry created by one user can be reused for other users asking semantically similar questions to the same agent. This leads to faster response times and more consistent answers across the user base.

The Cache Manager provides full visibility into how agentic caching is operating. It provides administrators the tools to fine‑tune and maintain the cache to meet performance and governance requirements. Use the Cache Manager to:

Inspect cache groups (buckets) and their questions.
Review and edit the SQL and other technical details behind cached responses.
Manage caching behavior such as enabling and disabling caching and controlling how new buckets are created.
Delete cache entries or clear the cache entirely for a given agent.

For steps to use the Cache Manager, see the following procedures:

Access the Cache Manager
Change Cache Modes
Delete All Caches
Create New Cache BucketsThis procedure walks you through and describes how cache buckets are created and updated.
Modify Cache Buckets

Using Agentic Caching

Combine agentic caching with normal question flows to balance performance, control, and flexibility.

Prompts

Prompts are always cached and their cache is reused, independent of the global cache mode. This guarantees maximum performance and consistency for these key, highly reused entry‑point questions. To create prompts, see Add Visualizations to Agents Using the Prompt Panel.

Reused vs Recomputed Caches

Cache reuse at the agent level depends on several factors. Even when a matching bucket exists, the system may need to recompute parts of the pipeline.

SQL, execution, answer, and visualization must be recalculated using cached SQL as reference in the following cases:

Dynamic time or dynamic user context

Questions involving expressions like last week, last year, today, this month, my manager, my team, me require recomputation because the time or user context changes over time or per user.
Data changes (cube refresh after cache date)

If the underlying cube's refresh time is later than the cache timestamp, previously cached executions may be outdated. The system re‑executes the SQL and regenerates the answer and visualization.
Same pattern, different parameters

When a question matches a bucket pattern but specific parameters differ (for example, a different region, product, or period), the cached SQL structure is reused as a template, but:
- Parameterized SQL is regenerated
- The query is re‑executed
- A new answer and visualization are produced

In all these situations, the cache still provides value by supplying the SQL blueprint and analysis structure, while ensuring that results remain current and contextually correct.

Access the Cache Manager

Requirements

To access the Cache Manager, you must have the Strategy AI: Administer Agents privilege.

Create a new bot or open an existing agent and click the Edit icon in the upper right.
In the toolbar, click the Cache Manager icon .

Change Cache Modes

Access the Cache Manager.
Click the Settings icon in the Caching Buckets pane.
From the Caching Mode drop-down list, select one of the following:
- Full Caching: Every question any user asks generates a cache entry. Questions can still form groups/buckets based on similarity.
- Existing Only: No new cache groups/buckets are created; only existing cache buckets are used. New questions are added to existing buckets if the questions are detected as similar by intent and columns used.
- Turn Off: Caching is disabled, preventing the generation of new caches.
  
  Prompts are always cached and always reuse the cache, even if caching is disabled.
Close the Cache Manager.
Save the agent.

Delete All Caches

Access the Cache Manager.
Click the Settings icon in the Caching Buckets pane.
Click Delete All Caches.
Close the Cache Manager.
Save the agent.

Create New Cache Buckets

This procedure walks you through and describes how cache buckets are created and updated.

Access the Cache Manager.
Click the Settings icon in the Caching Buckets pane.
From the Caching Mode drop-down list, select Full Caching. This ensures that new user questions can create new cache entries and buckets.
Close the Cache Manager.
Ask the agent a new question.
Access the Cache Manager.
A new bucket displays in the Caching Buckets pane. In the example below, the query "Identify the product sub-category with the highest total profit". The agent's interpretation of the query displays; in this case, it is "Identify the product sub-category with the highest summed profit". The objects used to answer the query are also provided; in this example they are Sub-Category_ID and Profit.
Click the bucket to open it. The pane on the right displays the questions the bucket contains, the generated SQL, and the attributes and metrics used to construct the answer.
To expand and review the full answer for a cache entry, click the right arrow next to the question.

The first question that created the bucket is treated as the parent question and is shown with a similarity score of 100. The example below shows only a single question.
To add more entries to this bucket, ask semantically similar questions in the agent chat. For example, paraphrase the original query by asking “Find product sub-category with highest profit.”. This is evaluated as similar and, if sufficiently close, added to the same bucket with its own similarity score relative to the parent.

In the image below, the blue cache icon indicates that the answer was generated using cached data. To regenerate the answer without using the cache, click the cache icon .

When such a semantically similar, paraphrased question is asked, it is automatically grouped into the same bucket as the original (parent) question in the Cache Manager. In the bucket view, this new question displays as a child entry with a similarity score indicating how closely it matches the parent question.
To create another bucket, ask a completely new question that is not semantically related to the previous ones. For example, ask the agent "What is the average order value in Central region during Q2 2014?".

Because this question represents a different intent, the caching mechanism creates a new bucket in the Cache Manager. In the Cache Manager shown below, a separate bucket corresponding to this question displays. It contains its own parent entry, generated SQL, and associated attributes and metrics.

Children of a bucket include not only paraphrased questions that reuse the exact same SQL, but also questions that result in slightly different SQL conditions while preserving the same overall intent and structure. For example, a question such as “What is the average order value in EMEA region during Q1 2014?” can be grouped into the same bucket as “What is the average order value in Central region during Q2 2014?”. In this case, the underlying query pattern (the average order value by region and time period) remains the same, but individual filter values (region, quarter, year) differ.

Modify Cache Buckets

A key advantage of agentic caching is that it gives users control over previously asked questions and their cached answers. If the agent returns an incorrect or sub‑optimal result, use the Cache Manager to adjust how that answer is produced. Specifically, you can:

Modify the attributes and metrics used for that cached entry, ensuring the logic matches your analytical requirements.
Edit the underlying SQL query used to retrieve the data by correcting filters, joins, or any other query details.
Merge cache buckets when you identify that multiple buckets represent the same analytical intent and should share a common cache group.
Move individual entries between buckets to regroup specific questions under a more appropriate parent bucket.

These changes allow you to refine existing cache entries so that future answers to the same or similar questions are accurate and aligned with your expectations.

Change the Columns Used

Changing columns (attributes and metrics) in cached entries is useful when the initial answer is structurally correct but not aggregated or detailed at the desired level.

For example, when the question “Where is lowest shipping cost on average?” is asked, the agent may respond with the country that has the lowest average shipping cost, while the required aggregation level is region.

In this case, use the Cache Manager to:

Open the corresponding bucket and cached entry.
Replace the Country attribute with Region, so that the query aggregates results at the regional level.

After this change is saved, subsequent executions of the same question “Where is lowest shipping cost on average?” reuse the updated cache entry and return results aggregated by Region.

Furthermore, attributes and metrics in the cached entry can be added or removed to better align the result set with the analytical intent. For example:

Add more metrics (such as total shipments or number of orders) to provide more context.
Remove irrelevant attributes to simplify the results.

By adjusting the attributes and metrics associated with cached entries, repeated and similar questions are answered in a way that best fits the intended analysis.

Ask the query in the agent. The example query “Where is lowest shipping cost on average?” is used, with an answer of Equatorial Guinea, a country, being provided.
Access the Cache Manager. The cached bucket uses Country_ID.
To fix the example issue, drag Country_ID from the bucket to the datasets panel. Drag Region_ID from the datasets panel to the bucket.
Close the Cache Manager.
In the agent, hover your cursor over the question and click Ask Again .

Now the answer is EMEA, a region instead of a country. Notice the blue cache icon indicating that the answer was generated using cached data.
Save the agent.

Edit the SQL Query

Modifying the SQL for cached entries provides full control over how answers are computed, beyond simple attribute or metric changes.

Continuing the previous example, after asking “Where is lowest shipping cost on average?”, the initial answer may aggregate by Country, while the required breakdown is by Region and State. Instead of only changing the attributes, you can edit the underlying SQL for the cached entry directly in the Cache Manager.

For example, the SQL can be modified to:

Change the grouping level from Country to Region and State.
Add or adjust filters, joins, or calculation logic.
Include additional columns or derived metrics that are not part of the original query.

After saving the updated SQL, subsequent executions of the same question reuse this modified query, so the answers reflect aggregation by Region and State (or any other logic defined in the SQL). In general, any valid change to the SQL can be applied so that the cached answer aligns precisely with reporting and analytical requirements.

Ask the query in the agent. The example query “Where is lowest shipping cost on average?” is used, with the answer of the EMEA region since the cache bucket was updated in the previous procedure.
Access the Cache Manager. The cached bucket uses Region_ID.
Select the bucket to display its details in the pane on the right.
Click View SQL. The SQL Query window opens.
To fix the example issue, type "State_ID" in both the SELECT statements, the GROUP BY statement, and the ORDER BY statement, as shown below.
In the toolbar, click Verify SQL to ensure your changes are valid. Check the SQL Output at the bottom to ensure the answer looks correct.
In the toolbar, click Save SQL.
After the successful save message displays, close the SQL Query window.
Close the Cache Manager.
In the agent, hover your cursor over the question and click Ask Again .

Now the answer is the Matabeland North region in the Africa region, instead of just a region. Notice the blue cache icon indicating that the answer was generated using cached data.
Save the agent.

Manage Buckets

Use the Cache Manager to restructure cache buckets so that they better reflect how questions are logically grouped.

Merge Buckets

Merge buckets when multiple buckets clearly represent the same analytical intent. Instead of maintaining several near‑duplicate groups, combine them into a single bucket so that all related questions share the same cache logic and are maintained in one place.

Access the Cache Manager.
Select the check box for each bucket to merge. Click the menu icon next to one of the selected buckets and select Merge Together.
The Merge Interpretation Buckets window displays. Select the bucket to merge into, that is, the bucket that keeps its name and properties. The other buckets are added to this target bucket.
You can review the SQL for a bucket by clicking View SQL in its row.

The SQL Query window opens. After you have reviewed the SQL, close the window.
Click Merge All.
Close the Cache Manager.
Save the agent.

Move Items Between Buckets

Moving an item (a specific question and its answer) from one bucket to another is useful when it was initially grouped incorrectly. This lets the cache be “re-filed” so that each question is contained in the bucket whose logic best matches its intent.

Access the Cache Manager.
Select the bucket to display its details in the pane on the right.
Select the item to move and do one of the following:
- Click the menu icon next to it, point to Move To, and select the bucket to move it to.
- Drag it to another bucket in the Caching Buckets pane.

The item is moved to the target bucket.

Close the Cache Manager.
Save the agent.

Create a New Bucket From an Item

Sometimes a question inside a bucket turns out to represent a distinct use case that should not share logic with the rest of the group. In that case, create a new bucket from that item (a specific question and its answer), effectively promoting it to the parent of its own cache group, with its own independently managed logic.

Access the Cache Manager.
Select the bucket to display its details in the pane on the right.
Select the item to move, click the menu icon next to it, point to Move To, and select Create New Bucket.

The item is removed from its original bucket.

The item is now contained in a new bucket, named for the item.
Close the Cache Manager.
Save the agent.

Regenerate Interpretations

When Custom Instructions (CI) are changed, existing cached items may need their interpretations to be regenerated. Similarity between questions is evaluated based on these interpretations, so if CI changes, the interpretation of a question can change as well. This affects bucket assignment and similarity scores.

The Regenerate Interpretations action updates interpretations for cached items under the current bucket according to the latest CI.

During regeneration, all child items in the bucket are removed and only the parent item is preserved. Subsequent user questions repopulate the bucket based on the new interpretations.

In the agent, change the custom instructions on the Advanced Settings panel.
- In the example below, the custom instructions have details such as changing average shipping cost to mean shipping cost and differentiating between revenue, profit, and sales.
- Remove the details, leaving only the introductory paragraph.
Save the agent.
Access the Cache Manager.
Select the check box for each bucket to regenerate. Click the menu icon next to one of the selected buckets and select Regenerate Interpretations. In the example below, all the buckets are selected. The "where is lowest shipping cost on average" is interpreted as the lowest mean shipping cost, as determined by the original custom instructions.
Review the buckets after the interpretation regeneration message displays. The "where is lowest shipping cost on average" is interpreted as the lowest average shipping cost, rather than the mean.
Close the Cache Manager.
Save the agent.