fix: disable think and keep alive until manual unload

agents/ollama.ts

@@ -117,11 +117,9 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
         model,
         messages,
         tools,
-        keep_alive: "10m",
-        options: {
-          think: true,
-          num_ctx: 262144,
-        } as Record<string, unknown>,
+        keep_alive: -1,
+        think: false,
+        options: { num_ctx: 262144 },
       });
     } catch (err) {
       const lastError = err instanceof Error ? err.message : String(err);
@@ -162,6 +160,16 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
       } else {
         log.info("» agent finished (no tool calls)");
       }
+
+      // unload the model from memory immediately since we're done with it, to free up resources for future runs. see #829.
+      try {
+        await ollama.generate({ model, keep_alive: 0, prompt: "" });
+        log.info(`» unloaded model ${model} from Ollama`);
+      } catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        log.warning(`» failed to unload model ${model}: ${msg}`);
+      }
+
       return {
         success: true,
         output: assistantMessage.content || undefined,
@@ -239,6 +247,16 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
   }
 
   log.warning(`» agent hit max iterations (${MAX_ITERATIONS})`);
+
+  // unload the model from memory immediately since we're done with it, to free up resources for future runs. see #829.
+  try {
+    await ollama.generate({ model, keep_alive: 0, prompt: "" });
+    log.info(`» unloaded model ${model} from Ollama`);
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    log.warning(`» failed to unload model ${model}: ${msg}`);
+  }
+
   return {
     success: false,
     error: `Agent exceeded maximum iterations (${MAX_ITERATIONS})`,