diff --git a/backend/src/services/ai.rs b/backend/src/services/ai.rs
index 3b68b19..b5ad1b0 100644
--- a/backend/src/services/ai.rs
+++ b/backend/src/services/ai.rs
@@ -14,6 +14,8 @@ use serde_json::{Value, json};
 use std::fs;
 use std::path::{Path, PathBuf};
 use std::sync::{Mutex, OnceLock};
+use std::thread;
+use std::time::{Duration, Instant};
 use uuid::Uuid;
 
 use crate::{
@@ -40,6 +42,8 @@ pub(crate) const REINDEX_EMBEDDING_BATCH_SIZE: usize = 4;
 const EMBEDDING_DIMENSION: usize = 384;
 const LOCAL_EMBEDDING_MODEL_LABEL: &str = "fastembed / local all-MiniLM-L6-v2";
 const LOCAL_EMBEDDING_CACHE_DIR: &str = "storage/ai_embedding_models/all-minilm-l6-v2";
+const LOCAL_EMBEDDING_IDLE_TIMEOUT_SECS: u64 = 300;
+const LOCAL_EMBEDDING_REAPER_INTERVAL_SECS: u64 = 30;
 const LOCAL_EMBEDDING_BASE_URL: &str =
     "https://huggingface.co/Qdrant/all-MiniLM-L6-v2-onnx/resolve/main";
 const LOCAL_EMBEDDING_FILES: [&str; 5] = [
@@ -50,7 +54,13 @@ const LOCAL_EMBEDDING_FILES: [&str; 5] = [
     "tokenizer_config.json",
 ];
 
-static TEXT_EMBEDDING_MODEL: OnceLock<Mutex<TextEmbedding>> = OnceLock::new();
+static TEXT_EMBEDDING_MODEL: OnceLock<Mutex<Option<LocalEmbeddingRuntime>>> = OnceLock::new();
+static TEXT_EMBEDDING_REAPER_STARTED: OnceLock<()> = OnceLock::new();
+
+struct LocalEmbeddingRuntime {
+    model: TextEmbedding,
+    last_used_at: Instant,
+}
 
 #[derive(Clone, Debug)]
 struct AiImageRuntimeSettings {
@@ -403,18 +413,78 @@ fn load_local_embedding_model() -> Result<TextEmbedding> {
         .map_err(|error| Error::BadRequest(format!("本地 embedding 模型初始化失败: {error}")))
 }
 
-fn local_embedding_engine() -> Result<&'static Mutex<TextEmbedding>> {
-    if let Some(model) = TEXT_EMBEDDING_MODEL.get() {
-        return Ok(model);
+fn local_embedding_state() -> &'static Mutex<Option<LocalEmbeddingRuntime>> {
+    TEXT_EMBEDDING_MODEL.get_or_init(|| Mutex::new(None))
+}
+
+fn ensure_local_embedding_reaper_started() {
+    TEXT_EMBEDDING_REAPER_STARTED.get_or_init(|| {
+        if let Err(error) = thread::Builder::new()
+            .name("local-embedding-reaper".to_string())
+            .spawn(|| {
+                let idle_timeout = Duration::from_secs(LOCAL_EMBEDDING_IDLE_TIMEOUT_SECS);
+                let check_interval = Duration::from_secs(LOCAL_EMBEDDING_REAPER_INTERVAL_SECS);
+
+                loop {
+                    thread::sleep(check_interval);
+
+                    let Some(state) = TEXT_EMBEDDING_MODEL.get() else {
+                        continue;
+                    };
+
+                    let mut guard = match state.lock() {
+                        Ok(guard) => guard,
+                        Err(_) => {
+                            tracing::warn!("failed to lock local embedding model for idle cleanup");
+                            continue;
+                        }
+                    };
+
+                    let should_unload = guard
+                        .as_ref()
+                        .is_some_and(|runtime| runtime.last_used_at.elapsed() >= idle_timeout);
+
+                    if should_unload {
+                        *guard = None;
+                        tracing::info!(
+                            "unloaded local embedding model after {} seconds of inactivity",
+                            LOCAL_EMBEDDING_IDLE_TIMEOUT_SECS
+                        );
+                    }
+                }
+            })
+        {
+            tracing::warn!("failed to start local embedding reaper thread: {error}");
+        }
+    });
+}
+
+fn with_local_embedding_engine<T>(
+    operation: impl FnOnce(&mut TextEmbedding) -> Result<T>,
+) -> Result<T> {
+    ensure_local_embedding_reaper_started();
+
+    let state = local_embedding_state();
+    let mut guard = state
+        .lock()
+        .map_err(|_| Error::BadRequest("本地 embedding 模型当前不可用，请稍后重试".to_string()))?;
+
+    if guard.is_none() {
+        tracing::info!("loading local embedding model into memory");
+        *guard = Some(LocalEmbeddingRuntime {
+            model: load_local_embedding_model()?,
+            last_used_at: Instant::now(),
+        });
     }
 
-    let model = load_local_embedding_model()?;
+    let runtime = guard
+        .as_mut()
+        .ok_or_else(|| Error::BadRequest("本地 embedding 模型未能成功缓存".to_string()))?;
+    runtime.last_used_at = Instant::now();
 
-    let _ = TEXT_EMBEDDING_MODEL.set(Mutex::new(model));
-
-    TEXT_EMBEDDING_MODEL
-        .get()
-        .ok_or_else(|| Error::BadRequest("本地 embedding 模型未能成功缓存".to_string()))
+    let result = operation(&mut runtime.model);
+    runtime.last_used_at = Instant::now();
+    result
 }
 
 fn vector_literal(embedding: &[f64]) -> Result<String> {
@@ -793,24 +863,21 @@ async fn embed_texts_locally_with_batch_size(
     batch_size: usize,
 ) -> Result<Vec<Vec<f64>>> {
     tokio::task::spawn_blocking(move || {
-        let model = local_embedding_engine()?;
         let prepared = inputs
             .iter()
             .map(|item| prepare_embedding_text(kind, item))
             .collect::<Vec<_>>();
 
-        let mut guard = model.lock().map_err(|_| {
-            Error::BadRequest("本地 embedding 模型当前不可用，请稍后重试".to_string())
-        })?;
+        with_local_embedding_engine(|model| {
+            let embeddings = model
+                .embed(prepared, Some(batch_size.max(1)))
+                .map_err(|error| Error::BadRequest(format!("本地 embedding 生成失败: {error}")))?;
 
-        let embeddings = guard
-            .embed(prepared, Some(batch_size.max(1)))
-            .map_err(|error| Error::BadRequest(format!("本地 embedding 生成失败: {error}")))?;
-
-        Ok(embeddings
-            .into_iter()
-            .map(|embedding| embedding.into_iter().map(f64::from).collect::<Vec<_>>())
-            .collect::<Vec<_>>())
+            Ok(embeddings
+                .into_iter()
+                .map(|embedding| embedding.into_iter().map(f64::from).collect::<Vec<_>>())
+                .collect::<Vec<_>>())
+        })
     })
     .await
     .map_err(|error| Error::BadRequest(format!("本地 embedding 任务执行失败: {error}")))?