@@ -155,7 +155,7 @@ const TOOL_SAVE_DESCRIPTION =
   "Call this when the user states a durable preference, a project decision is made, or a lesson is learned. " +
   "Merges into an existing entry of the same type when the title matches.";
 const TOOL_SEARCH_DESCRIPTION =
-  "Search the cross-session memory store. Use when you need past context: how a problem was solved, user preferences, project decisions. Substring-matches title/content/tags, and augments results with semantic (vector) recall + optional rerank when an embeddings provider is configured. Returns matching entries with source and timestamps.";
+  "Search the cross-session memory store for past context (solved problems, user preferences, project decisions). IMPORTANT: relevant cross-session memories are already injected into your context every turn — only call this when the injected memory block does NOT contain what you need, or you need to look something specific up. Substring-matches title/content/tags, and augments results with semantic (vector) recall + optional rerank when an embeddings provider is configured. Returns matching entries with source and timestamps.";
 const TOOL_LIST_DESCRIPTION =
   "List memory entries by type, high-importance first, then newest, paginated. Set include_archived=true to also list archived (hidden) entries so they can be located and restored.";
 const TOOL_GET_DESCRIPTION =
 
@@ -550,6 +550,17 @@ export const Config = z.object({
   // recall_runs 滚动清理保留天数。
   recallRetentionDays: z.natural().min(1).max(3650).default(90),
 
+  // --- tool exposure: 慢模型/轻量模型的工具往返节流（v0.8.6）----------------
+  // 跨会话记忆已由 inject.js 每轮自动注入系统提示词，memory_search 只用于
+  // 「注入里没有、需要深挖」的补充检索；memory_archive 是隐藏/恢复条目的整理
+  // 操作，正常会话里很少需要。轻量模型对「何时该用工具」判断弱，容易每轮
+  // 顺手调一遍——每次工具调用都是一次串行往返（生成参数→执行→回填→再生成），
+  // 在慢模型上会被放大成明显卡顿。这两个开关允许直接隐藏对应工具（默认全
+  // 开=行为不变），隐藏后模型根本看不到它，也就不会调。也走 feature_flags
+  // 白名单，面板可启停=线上回滚开关。
+  disableMemorySearch: z.boolean().default(false),
+  disableMemoryArchive: z.boolean().default(false),
+
   // --- scope: v0.8.0 A1 存储层（issue #17）--------------------------------
   // 总开关默认关：关闭时写入不标注 scope、去重维持 (type, title) 现状，行为
   // 逐字节不变。开启后 memory_save 写入 agent_scope（session header 的
 
@@ -85,6 +85,10 @@ const FEATURE_FLAG_BOOLEANS = [
   // Issue #17（v0.8.0 A3）：strictScope 硬过滤——他 scope 完全不可见（关闭时
   // 为 A2 软隔离：降权保留可见）。
   "strictScope",
+  // 工具暴露开关（v0.8.6）：记忆已每轮自动注入，memory_search/memory_archive
+  // 在慢/轻量模型上是多余往返，面板可关（默认关=行为不变）。
+  "disableMemorySearch",
+  "disableMemoryArchive",
   // 嵌套对象开关：config.js 里是 memoryQualityFilter / llmAudit 对象的 enabled
   // 子字段。kv 按点号键平铺存（"memoryQualityFilter.enabled": false），index.js
   // 合并时展开回嵌套对象，api.js 的 effective 从对象子字段取值。
 
@@ -181,7 +181,7 @@ export function createTools(ctx, service, config, embedder) {
       name: "memory_search",
       description: withToolGuide(
         "memory_search",
-        "Search the cross-session memory store. Use when you need past context: how a problem was solved, user preferences, project decisions. Substring-matches title/content/tags, and augments results with semantic (vector) recall + optional rerank when an embeddings provider is configured. Returns matching entries with source and timestamps."
+        "Search the cross-session memory store for past context (solved problems, user preferences, project decisions). IMPORTANT: relevant cross-session memories are already injected into your context every turn — only call this when the injected memory block does NOT contain what you need, or you need to look something specific up. Substring-matches title/content/tags, and augments results with semantic (vector) recall + optional rerank when an embeddings provider is configured. Returns matching entries with source and timestamps."
       ),
       // A2 检索接线：occurred 时间窗 + 会话 scope 加成（见 execute）。
       parameters: {
@@ -435,7 +435,8 @@ export function createTools(ctx, service, config, embedder) {
       description:
         "Archive a memory (hide it from active lists, search, injection and dream consolidation) or restore it. " +
         "Archived entries stay in storage and are recoverable: pass archived=false to restore, and use memory_list with " +
-        "include_archived=true to find archived entries.",
+        "include_archived=true to find archived entries. " +
+        "Only archive when the user asks, or an entry is clearly stale/no longer relevant — do not archive proactively mid-session to tidy up.",
       parameters: {
         id: { type: "string", required: true, description: "Memory id" },
         archived: { type: "boolean", description: "Archive (true, default) or restore (false) the entry" }
@@ -676,8 +677,20 @@ export function createTools(ctx, service, config, embedder) {
     })
   ];
 
+  // Tool-exposure gate (v0.8.6): 跨会话记忆已自动注入，memory_search /
+  // memory_archive 在慢/轻量模型上是多余往返，面板开关可直接隐藏工具——
+  // 模型看不到就不会调，比在描述里劝它少调用更可靠。注意：live patch reload
+  // 下已注册的工具不会被宿主反注册，排除只对全新会话生效。
+  const hiddenTools = new Set();
+  if (config?.disableMemorySearch === true) hiddenTools.add("memory_search");
+  if (config?.disableMemoryArchive === true) hiddenTools.add("memory_archive");
+
   for (const tool of tools) {
-          if (registeredTools.has(tool.name)) {
+    if (hiddenTools.has(tool.name)) {
+      ctx.logger?.info?.(`[dsh-mneme] tool "${tool.name}" hidden by config (disableMemorySearch/disableMemoryArchive)`);
+      continue;
+    }
+    if (registeredTools.has(tool.name)) {
         ctx.logger?.warn?.(`[dsh-mneme] tool "${tool.name}" already registered, skipping duplicate`);
         continue;
       }
 
@@ -550,6 +550,17 @@ export const Config = z.object({
   // recall_runs 滚动清理保留天数。
   recallRetentionDays: z.natural().min(1).max(3650).default(90),
 
+  // --- tool exposure: 慢模型/轻量模型的工具往返节流（v0.8.6）----------------
+  // 跨会话记忆已由 inject.js 每轮自动注入系统提示词，memory_search 只用于
+  // 「注入里没有、需要深挖」的补充检索；memory_archive 是隐藏/恢复条目的整理
+  // 操作，正常会话里很少需要。轻量模型对「何时该用工具」判断弱，容易每轮
+  // 顺手调一遍——每次工具调用都是一次串行往返（生成参数→执行→回填→再生成），
+  // 在慢模型上会被放大成明显卡顿。这两个开关允许直接隐藏对应工具（默认全
+  // 开=行为不变），隐藏后模型根本看不到它，也就不会调。也走 feature_flags
+  // 白名单，面板可启停=线上回滚开关。
+  disableMemorySearch: z.boolean().default(false),
+  disableMemoryArchive: z.boolean().default(false),
+
   // --- scope: v0.8.0 A1 存储层（issue #17）--------------------------------
   // 总开关默认关：关闭时写入不标注 scope、去重维持 (type, title) 现状，行为
   // 逐字节不变。开启后 memory_save 写入 agent_scope（session header 的
 
@@ -85,6 +85,10 @@ const FEATURE_FLAG_BOOLEANS = [
   // Issue #17（v0.8.0 A3）：strictScope 硬过滤——他 scope 完全不可见（关闭时
   // 为 A2 软隔离：降权保留可见）。
   "strictScope",
+  // 工具暴露开关（v0.8.6）：记忆已每轮自动注入，memory_search/memory_archive
+  // 在慢/轻量模型上是多余往返，面板可关（默认关=行为不变）。
+  "disableMemorySearch",
+  "disableMemoryArchive",
   // 嵌套对象开关：config.js 里是 memoryQualityFilter / llmAudit 对象的 enabled
   // 子字段。kv 按点号键平铺存（"memoryQualityFilter.enabled": false），index.js
   // 合并时展开回嵌套对象，api.js 的 effective 从对象子字段取值。
 
@@ -181,7 +181,7 @@ export function createTools(ctx, service, config, embedder) {
       name: "memory_search",
       description: withToolGuide(
         "memory_search",
-        "Search the cross-session memory store. Use when you need past context: how a problem was solved, user preferences, project decisions. Substring-matches title/content/tags, and augments results with semantic (vector) recall + optional rerank when an embeddings provider is configured. Returns matching entries with source and timestamps."
+        "Search the cross-session memory store for past context (solved problems, user preferences, project decisions). IMPORTANT: relevant cross-session memories are already injected into your context every turn — only call this when the injected memory block does NOT contain what you need, or you need to look something specific up. Substring-matches title/content/tags, and augments results with semantic (vector) recall + optional rerank when an embeddings provider is configured. Returns matching entries with source and timestamps."
       ),
       // A2 检索接线：occurred 时间窗 + 会话 scope 加成（见 execute）。
       parameters: {
@@ -435,7 +435,8 @@ export function createTools(ctx, service, config, embedder) {
       description:
         "Archive a memory (hide it from active lists, search, injection and dream consolidation) or restore it. " +
         "Archived entries stay in storage and are recoverable: pass archived=false to restore, and use memory_list with " +
-        "include_archived=true to find archived entries.",
+        "include_archived=true to find archived entries. " +
+        "Only archive when the user asks, or an entry is clearly stale/no longer relevant — do not archive proactively mid-session to tidy up.",
       parameters: {
         id: { type: "string", required: true, description: "Memory id" },
         archived: { type: "boolean", description: "Archive (true, default) or restore (false) the entry" }
@@ -676,8 +677,20 @@ export function createTools(ctx, service, config, embedder) {
     })
   ];
 
+  // Tool-exposure gate (v0.8.6): 跨会话记忆已自动注入，memory_search /
+  // memory_archive 在慢/轻量模型上是多余往返，面板开关可直接隐藏工具——
+  // 模型看不到就不会调，比在描述里劝它少调用更可靠。注意：live patch reload
+  // 下已注册的工具不会被宿主反注册，排除只对全新会话生效。
+  const hiddenTools = new Set();
+  if (config?.disableMemorySearch === true) hiddenTools.add("memory_search");
+  if (config?.disableMemoryArchive === true) hiddenTools.add("memory_archive");
+
   for (const tool of tools) {
-          if (registeredTools.has(tool.name)) {
+    if (hiddenTools.has(tool.name)) {
+      ctx.logger?.info?.(`[dsh-mneme] tool "${tool.name}" hidden by config (disableMemorySearch/disableMemoryArchive)`);
+      continue;
+    }
+    if (registeredTools.has(tool.name)) {
         ctx.logger?.warn?.(`[dsh-mneme] tool "${tool.name}" already registered, skipping duplicate`);
         continue;
       }
 
@@ -663,8 +663,8 @@ test("GET /api/dsh-mneme/features returns empty overrides and effective config d
   // issue #257 新增 sleepMaxTokens、issue #258 新增总览路由两键与
   // dreamSummaryMaxInputs、issue #230 新增 documentMemoryEnabled/
   // documentInjectBudget、issue #249 第一批新增 injectGuidanceEnabled/
-  // pinnedInjectBudget）
-  assert.equal(Object.keys(data.effective).length, 59 + 3 + 2 + 1 + 2 + 2);
+  // pinnedInjectBudget，v0.8.6 新增 disableMemorySearch/disableMemoryArchive）
+  assert.equal(Object.keys(data.effective).length, 59 + 3 + 2 + 1 + 2 + 2 + 2);
   assert.equal(data.effective.dreamSkipInvalid, true);
   assert.equal(data.effective.allowCrossTypeMerge, false);
   assert.equal(data.effective.dreamMinIntervalMinutes, 0);
 
@@ -54,6 +54,38 @@ test("registers ten tools with correct names", () => {
   assert.deepEqual(names, ["memory_archive", "memory_delete", "memory_forget", "memory_get", "memory_list", "memory_register_document", "memory_runtime", "memory_save", "memory_search", "memory_update"]);
 });
 
+// v0.8.6 tool-exposure gate: disabling a tool hides it from the model entirely
+// (it is not registered), which is the reliable lever against over-calling on
+// slow/lightweight models — the injected memory block already covers recall.
+test("disableMemorySearch hides only memory_search", () => {
+  const { registered } = setup({}, { disableMemorySearch: true });
+  const names = registered.map((t) => t.name).sort();
+  assert.deepEqual(names, ["memory_archive", "memory_delete", "memory_forget", "memory_get", "memory_list", "memory_register_document", "memory_runtime", "memory_save", "memory_update"]);
+});
+
+test("disableMemoryArchive hides only memory_archive", () => {
+  const { registered } = setup({}, { disableMemoryArchive: true });
+  const names = registered.map((t) => t.name).sort();
+  assert.deepEqual(names, ["memory_delete", "memory_forget", "memory_get", "memory_list", "memory_register_document", "memory_runtime", "memory_save", "memory_search", "memory_update"]);
+});
+
+test("both disable flags hide both tools", () => {
+  const { registered } = setup({}, { disableMemorySearch: true, disableMemoryArchive: true });
+  const names = registered.map((t) => t.name).sort();
+  assert.deepEqual(names, ["memory_delete", "memory_forget", "memory_get", "memory_list", "memory_register_document", "memory_runtime", "memory_save", "memory_update"]);
+});
+
+// Description discipline: with all tools exposed, the descriptions must steer a
+// weak model away from calling them every turn.
+test("memory_search/archive descriptions carry tool-use discipline", () => {
+  const { registered } = setup();
+  const search = registered.find((t) => t.name === "memory_search");
+  const archive = registered.find((t) => t.name === "memory_archive");
+  assert.match(search.description, /already injected into your context every turn/i);
+  assert.match(search.description, /only call this when/i);
+  assert.match(archive.description, /do not archive proactively mid-session/i);
+});
+
 test("compiled schemas pass the enforced DSH subset (defineTool projection)", () => {
   const { registered } = setup();
   assert.equal(registered.length, 10);