diff --git a/openbot/docs/configuration.md b/openbot/docs/configuration.md
index b2c5113d..0294c8eb 100644
--- a/openbot/docs/configuration.md
+++ b/openbot/docs/configuration.md
@@ -239,6 +239,8 @@ Most gateways publish a model list, which is the way to check a name before conf
 
 Two things are worth knowing before pointing a deployment at any gateway. Not every catalogue entry accepts tools, and a Bot without tool calling cannot drive its computer; the model list says which do. And `BOT_RESPONSES_API=true` needs an endpoint that implements the Responses API, not only chat completions.
 
+With `OPENAI_BASE_URL` set, the built-in Bot sends its Responses API requests with `store: false` and each earlier step of a turn in full, because an endpoint other than OpenAI keeps no stored responses to refer back to. A reasoning model's thinking is therefore not carried from one step of a turn to the next, including on a gateway in front of OpenAI itself.
+
 ## Voice dictation
 
 Dictation inserts recorded speech into the composer for review before sending. It works with any
diff --git a/openbot/server/src/copilot.ts b/openbot/server/src/copilot.ts
index ba0c6359..a4defe39 100644
--- a/openbot/server/src/copilot.ts
+++ b/openbot/server/src/copilot.ts
@@ -469,6 +469,17 @@ export function builtInAgentConfiguration(
       ...(learnedSkills?.catalog ? [learnedSkills.catalog] : []),
     ].join("\n\n"),
     ...(planModel ? {} : { apiKey: apiKey ?? undefined }),
+    /*
+     * An OpenAI-compatible endpoint other than OpenAI (OPENAI_BASE_URL) keeps no response store, so
+     * the Responses API must resend earlier items in full. With `store` on, the AI SDK sends a
+     * reasoning model's previous thinking back as `item_reference`, and an endpoint such as Ollama
+     * refuses the whole step after the first tool call ("unknown input item type").
+     */
+    ...(!planModel &&
+    model.provider === "openai" &&
+    process.env.OPENAI_BASE_URL?.trim()
+      ? { providerOptions: { openai: { store: false } } }
+      : {}),
     /*
      * A run stops after one step unless told otherwise, which for a Bot with tools means it calls
      * one and never speaks: the tool executes, the result arrives, and the run ends before the model
diff --git a/openbot/server/tests/copilot.test.ts b/openbot/server/tests/copilot.test.ts
index 978a0811..6215a30f 100644
--- a/openbot/server/tests/copilot.test.ts
+++ b/openbot/server/tests/copilot.test.ts
@@ -219,6 +219,90 @@ describe("deployment model selection", () => {
     },
   );
 
+  /*
+   * A compatible endpoint keeps no response store. With `store` on, the AI SDK sends a step's
+   * reasoning back as an `item_reference`, which such an endpoint (Ollama, for one) refuses, so a
+   * reasoning model failed every turn at its first tool call.
+   */
+  test("a compatible endpoint is sent the earlier step in full, not by reference", async () => {
+    const recorder = new LLMock();
+    const originalBase = process.env.OPENAI_BASE_URL;
+    // The journal keeps a normalised body, so the request is read as sent, on its way through.
+    const bodies: { store?: unknown; input?: unknown }[] = [];
+    let proxy: ReturnType<typeof Bun.serve> | undefined;
+    try {
+      const base = await recorder.start();
+      proxy = Bun.serve({
+        hostname: "127.0.0.1",
+        port: 0,
+        async fetch(request) {
+          const body = await request.text();
+          bodies.push(JSON.parse(body));
+          return fetch(`${base}${new URL(request.url).pathname}`, {
+            method: "POST",
+            headers: request.headers,
+            body,
+          });
+        },
+      });
+      process.env.OPENAI_BASE_URL = `${proxy.url.origin}/v1`;
+      recorder.on({ hasToolResult: true }, { content: "The balance is 42." });
+      recorder.onMessage(/balance/, {
+        reasoning: "Read the balance before answering.",
+        toolCalls: [
+          {
+            id: "balance-call",
+            name: "read_balance",
+            arguments: { account: "demo" },
+          },
+        ],
+      });
+      const agents = await resolveRuntimeAgents(
+        async () => [
+          {
+            id: "general-assistant",
+            name: "General Assistant",
+            type: "built_in" as const,
+            systemPrompt: "Answer using the granted balance tool.",
+          },
+        ],
+        { provider: "openai", defaultModel: "selected-local-model" },
+        async () => "synthetic-model-key",
+        undefined,
+        async () => [
+          {
+            name: "read_balance",
+            ref: "bank/read_balance",
+            description: "Read the balance.",
+            parameters: z.object({ account: z.string() }),
+            execute: async () => "42",
+          },
+        ],
+      );
+      const agent = agents["general-assistant"]?.clone();
+      if (!agent) throw new Error("Expected General Assistant.");
+      agent.addMessage({
+        id: "compatible-store-request",
+        role: "user",
+        content: "Read my balance.",
+      });
+      await agent.runAgent();
+
+      expect(agent.messages.at(-1)).toMatchObject({
+        role: "assistant",
+        content: "The balance is 42.",
+      });
+      expect(bodies).toHaveLength(2);
+      expect(bodies.map((body) => body.store)).toEqual([false, false]);
+      expect(JSON.stringify(bodies[1]?.input)).not.toContain("item_reference");
+    } finally {
+      if (originalBase === undefined) delete process.env.OPENAI_BASE_URL;
+      else process.env.OPENAI_BASE_URL = originalBase;
+      await proxy?.stop(true);
+      await recorder.stop();
+    }
+  });
+
   /*
    * Switching the provider away from the package's own, without naming a model. The two fallbacks are
    * not the same source: OpenAI's is the built-in Bot's own (it speaks the Responses API), Anthropic's