@Tool tool is called endlessly with chat memory enabled when using M8

[joshlong]

https://github.com/joshlong-attic/2026-05-29-infinite-tool-call-in-m8/blob/main/README.md

that is a reproducer

there's an oddity with infinite tool calling in Spring AI M8 + memory

you need ollama to make this run.

im using some big models but you might be able to get away with less.

them visit:

http://127.0.0.1:8080/josh/ask?question=fantastic.%20when%20can%20I%20pick%20up%20Prancer%20(dog%20ID%2045)%20from%20the%20Lisbon%20Pooch%20Palace%20location%3F&continue

it calls the tool over and over. if u change the DB to use PostgreSQL or somethign youc an see that its adding many blank rows in the chat memory table, too

[stohirov]

I guess it is an interaction bug between ToolCallAdvisor and MessageChatMemoryAdvisor.
@tzolov What do you think about putting MessageChatMemoryAdvisor at a lower order number than a ToolCallAdvisor? so, it wraps outside the loop and runs once per turn.

[ThomasVitale]

I tried to run the reproducer using a smaller model (ministral-3:8b) and the tool seems to be called only once. Unfortunately, I cannot try with gemma4:26b due to memory constraints on my current machine.

This is the trace I see from the interaction:

And I also verified that I only get a single log message Scheduled 45 for 2026-06-02T07:08:45.667734Z which is printed whenever the tool is called.

I also tried replacing H2 with PostgreSQL to inspect the content of the table and it looks to capture correctly the conversation. Though, the intermediate messages related to the tool calling are blank, which I guess makes sense since there's no content, but it is odd to have blank messages in the memory. Perhaps something to consider as part of the new Memory APIs coming in 2.1.

+---------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+---------+--------------------------+
|conversation_id|content                                                                                                                                                                                                                                                                                                                              |type     |timestamp                 |
+---------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+---------+--------------------------+
|josh           |fantastic. when can I pick up Prancer (dog ID 45) from the Lisbon Pooch Palace location?                                                                                                                                                                                                                                             |USER     |2026-05-30 09:08:48.000000|
|josh           |                                                                                                                                                                                                                                                                                                                                     |ASSISTANT|2026-05-30 09:08:49.000000|
|josh           |                                                                                                                                                                                                                                                                                                                                     |TOOL     |2026-05-30 09:08:50.000000|
|josh           |You can pick up **Prancer** (Dog ID 45) from the **Lisbon Pooch Palace** location on **June 2, 2026**, at any time during our operating hours. The location is typically open from **9:00 AM to 6:00 PM** (local time).                                                                                                              |ASSISTANT|2026-05-30 09:08:51.000000|
|               |Would you like assistance with anything else, such as transportation or additional details about Prancer? 😊                                                                                                                                                                                                                         |         |                          |
+---------------+-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+---------+--------------------------+

@stohirov the MessageChatMemoryAdvisor is inside the ToolCallingAdvisor to ensure we capture all the intermediate messages exchanged as part of tool calling, solving the issue in Spring AI 1.x were they were basically "invisible".

[stohirov]

@ThomasVitale ohhh, that makes sense. I will try to reproduce this issue with a larger model a bit later as well (gemma4:26b) cus of memory constraints on my current laptop.

What I am thinking hypothetically is that ToolCallAdvisor rebuilds instructions from toolExecutionResult.conversationHistory() which includes system, user, assistant tool call and a tool response and then, MessageChatMemoryAdvisor.before() also prepends what it already saved ([user, assistant-toolcall]). So the prompt sent on iteration 2 carries a duplicated user message a duplicated blank assistant-tool-call message.

I will come back to comment on this a bit later.

[ThomasVitale]

Ok, I managed to run the example with gemma4:26b and in that case I see the infinite tool calling. So it seems to be something specific to the model used. I tried with a few different models from the Mistral and Granite families, and for those the execution happens correctly.

This is what I'm running, in case you'd like to run the same (I configured the observability part on top of Josh's app): https://github.com/ThomasVitale/2026-05-29-infinite-tool-call-in-m8

[goutamadwant]

Thanks @ThomasVitale.. your trace helped narrow this a bit.

My read is that we should keep the advisor ordering as-is, since MessageChatMemoryAdvisor still needs to run inside the tool-calling loop to capture the intermediate assistant/tool messages.

I opened #6228 with narrower fix in MessageChatMemoryAdvisor.before().. When the next tool-call prompt is rebuilt from toolExecutionResult.conversationHistory() the memory advisor now avoids prepending stored memory and if those messages are already present in the prompt... also covered the MessageWindowChatMemory case where older messages may be evicted, so the stored memory is not necessarily at the beginning of the prompt.

Does this approach make sense to you?