Invalid parameter: messages with role 'tool' must be a response to a preceeding message with 'tool_calls'

Questions
1

Hi there! I’ve been having issues with something, I don’t know what it is. In the error message appears as “Bad request - please check your parametersInvalid parameter: messages with role ‘tool’ must be a response to a preceeding message with ‘tool_calls’.” Has something similar happend to you, I’m very stuck with this

Bad request - please check your parameters
Invalid parameter: messages with role ‘tool’ must be a response to a preceeding message with ‘tool_calls’.

Information on your n8n setup

  • n8n version: 2.30.6 (Self Hosted)
  • Database : PostgreSQL
2

Hi @Christian_G Welcome!
This happens when the chat history sent to the model has a tool message that isn’t directly preceded by an assistant message carrying its tool_calls, so the provider rejects the request. Your stored memory has gone out of sync for a session, usually from a run cut off mid tool call or a model swap, and since you’re on Postgres Chat Memory that broken history is saved and replayed on every run.
Clear that session’s memory. Quickest check: change the Session Key (or pass a new sessionId) and run again, the error should clear. To wipe the current session, connect a Chat Memory Manager node to the same Postgres Chat Memory, set Operation to Delete Messages and Delete Mode to All Messages, and run it once.

3

Hi Mr. Anshul, we will apply your recomendations an we’ll tell you how it works.
Thank you so much for your quick response​:folded_hands:.

4

Inspect the message array at the failed execution and pair each tool response with the assistant tool call that created it. Filtering or merging chat history can leave a tool message orphaned even when the current prompt looks fine.

5

Building on @Anshul_Namdev 's point: the root cause of this error in n8n AI Agents is usually **Window Buffer Memory Truncation**.

OpenAI’s API strictly requires that any message with `role: “tool”` MUST immediately follow an `assistant` message containing the matching `tool_calls` ID array.

Here is how this happens in n8n and how to prevent it permanently:

### The Root Cause:

If you use a **Window Buffer Memory** (or Postgres Memory) set to keep e.g. the last `10` messages:

1. An assistant message generates a `tool_call`.

2. The tool node executes and appends a `tool` response message.

3. As conversation progresses and reaches 11 messages, the Memory node drops the oldest message.

4. If the truncation split cuts off the `assistant` message with `tool_calls` but leaves the `tool` response message at the top of the history window, the payload sent to OpenAI becomes invalid.

### How to Fix It:

1. **Use `Window Buffer Memory` with Even Message Counts (or Increase K):**

Ensure your memory buffer size `K` is large enough so active tool-call pairs aren’t truncated mid-session.

2. **Clear Stale Sessions via `Chat Memory Manager`:**

If a session gets stuck with this error, use a **Chat Memory Manager Node** set to `Delete Messages` for that specific `sessionId`.

3. **Handle Tool Failures (Error Trigger / Sub-Workflow):**

If an MCP or Custom Tool fails mid-execution, wrap the tool logic in a `Try/Catch` or `Continue on Fail` setting so it always returns a valid JSON string payload to the agent instead of dropping the socket and corrupting the memory history.