diff --git a/agentmain.py b/agentmain.py
index e29e5b7e..1cac0ed9 100644
--- a/agentmain.py
+++ b/agentmain.py
@@ -49,6 +49,7 @@ def __init__(self):
         self.is_running = False; self.stop_sig = False
         self.llm_no = 0;  self.inc_out = False; self.verbose = True
         self.peer_hint = True
+        self._recall_done = False
         self.load_llm_sessions()
 
     def load_llm_sessions(self):
@@ -134,6 +135,15 @@ def run(self):
             self.is_running = True
             rquery = smart_format(raw_query.replace('\n', ' '), max_str_len=200)
             self.history.append(f"[USER]: {rquery}")
+            # --- startup recall: inject recent session context on first user message ---
+            if not self._recall_done and source == 'user':
+                self._recall_done = True
+                from frontends.startup_recall import build_startup_recall, is_trivial_message
+                if not is_trivial_message(raw_query):
+                    recall = build_startup_recall()
+                    if recall:
+                        raw_query = f"{recall}\n\n---\n[当前用户消息]\n{raw_query}"
+                        print(f'[GA] startup_recall injected ({len(recall)} chars)')
             
             sys_prompt = get_system_prompt() + getattr(self.llmclient.backend, 'extra_sys_prompt', '')
             if self.peer_hint: sys_prompt += f"\n[Peer] 用户提及其他会话/后台任务状态时: temp/model_responses/ (只找近期修改的文件尾部)\n"
diff --git a/frontends/startup_recall.py b/frontends/startup_recall.py
new file mode 100644
index 00000000..e74844ff
--- /dev/null
+++ b/frontends/startup_recall.py
@@ -0,0 +1,212 @@
+"""
+Startup Recall - Cold-start recent session memory injection.
+
+After a GA restart, automatically injects summaries of recent sessions
+into the first user message so the agent doesn't lose all context.
+
+Config via environment variables:
+    GA_STARTUP_RECALL_SESSIONS=3     Number of recent sessions to load (0 to disable)
+    GA_STARTUP_RECALL_LINES=6        Max lines per session to extract
+"""
+
+import ast
+import glob
+import json
+import os
+import re
+import time
+
+PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+RESTORE_GLOBS = (
+    os.path.join(PROJECT_ROOT, "temp", "model_responses", "model_responses_*.txt"),
+    os.path.join(PROJECT_ROOT, "temp", "model_responses_*.txt"),
+)
+RESTORE_BLOCK_RE = re.compile(
+    r"^=== (Prompt|Response) ===.*?\n(.*?)(?=^=== (?:Prompt|Response) ===|\Z)",
+    re.DOTALL | re.MULTILINE,
+)
+SUMMARY_RE = re.compile(r"<summary>\s*(.*?)\s*</summary>", re.DOTALL)
+
+# Lightweight messages that should NOT trigger recall injection
+_TRIVIAL_PATTERNS = re.compile(
+    r"^(hi|hello|hey|你好|您好|嗨|哈喽|在吗|在不在|\?|？|test|测试|ping|ok|好的|嗯|哦|谢谢|thx|thanks|ty)\s*$",
+    re.IGNORECASE,
+)
+
+# Patterns that indicate internal/system messages, not real user input
+_SKIP_PATTERNS = (
+    '[AUTO]', '用户已经离开', 'WORKING MEMORY', '[SYSTEM',
+    'FILE:filepath', '[DANGER]', 'cwd = /root', '### 用户当前消息',
+    '[CONSTITUTION]', 'show_linenos', 'MEMORY]', 'META-SOP',
+)
+
+
+def is_trivial_message(text: str) -> bool:
+    """Check if a message is too lightweight to trigger recall injection."""
+    return bool(_TRIVIAL_PATTERNS.match((text or "").strip()))
+
+
+def _restore_log_files():
+    files = []
+    for pattern in RESTORE_GLOBS:
+        files.extend(glob.glob(pattern))
+    return sorted(set(files))
+
+
+def _extract_user_facing_text(response_body: str) -> str:
+    """Extract user-visible text from a Response block.
+    
+    Parses the Python list of blocks, finds text blocks,
+    strips <summary> tags, and returns the user-facing content.
+    """
+    text = (response_body or "").strip()
+    try:
+        blocks = ast.literal_eval(text)
+        if isinstance(blocks, list):
+            for block in blocks:
+                if isinstance(block, dict) and block.get('type') == 'text':
+                    t = block.get('text', '')
+                    if not t:
+                        continue
+                    # Strip summary tags, keep user-facing part
+                    cleaned = re.sub(r'<summary>.*?</summary>', '', t, flags=re.DOTALL).strip()
+                    if cleaned and not cleaned.startswith('!!!Error'):
+                        return cleaned[:300]
+    except Exception:
+        pass
+    return ""
+
+
+def _extract_real_user_text(prompt_body: str) -> str:
+    """Extract the real user message from a Prompt block.
+    
+    Returns empty string for system/internal/AUTO messages.
+    """
+    text = (prompt_body or "").strip()
+    try:
+        obj = ast.literal_eval(text)
+        if not isinstance(obj, dict) or obj.get('role') != 'user':
+            return ""
+        content_list = obj.get('content', [])
+        if not isinstance(content_list, list):
+            return ""
+        for block in content_list:
+            if isinstance(block, dict) and block.get('type') == 'text':
+                t = block.get('text', '').strip()
+                if not t or len(t) < 5:
+                    continue
+                # Skip internal/system messages
+                if any(skip in t for skip in _SKIP_PATTERNS):
+                    continue
+                return t[:200]
+    except Exception:
+        pass
+    return ""
+
+
+def _extract_conversations(content: str, max_pairs: int) -> list:
+    """Extract real user-agent conversation pairs from a session file.
+    
+    Returns list of (user_text, agent_text) tuples.
+    """
+    blocks = RESTORE_BLOCK_RE.findall(content or "")
+    if not blocks:
+        return []
+
+    pairs = []
+    pending_prompt = None
+    for label, body in blocks:
+        if label == "Prompt":
+            pending_prompt = body
+        elif label == "Response" and pending_prompt is not None:
+            user_text = _extract_real_user_text(pending_prompt)
+            if user_text:
+                agent_text = _extract_user_facing_text(body)
+                if agent_text:
+                    pairs.append((user_text, agent_text))
+            pending_prompt = None
+
+    return pairs[-max_pairs:] if len(pairs) > max_pairs else pairs
+
+
+def _has_real_conversations(content: str) -> bool:
+    """Quick check if a session file has real user-agent conversations."""
+    blocks = RESTORE_BLOCK_RE.findall(content[:100000] or "")
+    prompt_bodies = [body for label, body in blocks if label == "Prompt"]
+    for p in prompt_bodies[-20:]:  # Check last 20 prompts
+        if _extract_real_user_text(p):
+            return True
+    return False
+
+
+def build_startup_recall(
+    max_sessions: int = None,
+    max_lines_per_session: int = None,
+) -> str:
+    """
+    Build a formatted recall string from recent session logs.
+
+    Returns empty string if:
+    - Feature is disabled (GA_STARTUP_RECALL=0)
+    - No session files found
+    - No extractable content
+    """
+    if max_sessions is None:
+        max_sessions = int(os.environ.get("GA_STARTUP_RECALL_SESSIONS", "3"))
+    if max_sessions <= 0:
+        return ""
+    if max_lines_per_session is None:
+        max_lines_per_session = int(os.environ.get("GA_STARTUP_RECALL_LINES", "6"))
+
+    files = _restore_log_files()
+    if not files:
+        return ""
+
+    # Sort by modification time, newest first
+    files.sort(key=lambda f: os.path.getmtime(f), reverse=True)
+
+    # Prefer files with real user interactions
+    interactive_files = []
+    for fpath in files[:20]:  # Check top 20 most recent
+        try:
+            with open(fpath, "r", encoding="utf-8", errors="replace") as f:
+                head = f.read(100000)
+            if _has_real_conversations(head):
+                interactive_files.append(fpath)
+        except Exception:
+            continue
+
+    candidates = interactive_files if interactive_files else files
+    recent_files = candidates[:max_sessions]
+
+    sessions = []
+    for fpath in recent_files:
+        try:
+            with open(fpath, "r", encoding="utf-8", errors="replace") as f:
+                content = f.read()
+        except Exception:
+            continue
+
+        conversations = _extract_conversations(content, max_lines_per_session)
+        if not conversations:
+            continue
+
+        mtime = os.path.getmtime(fpath)
+        ts = time.strftime("%Y-%m-%d %H:%M", time.localtime(mtime))
+
+        lines = []
+        for user_text, agent_text in conversations:
+            lines.append(f"[USER]: {user_text[:150]}")
+            lines.append(f"[Agent]: {agent_text[:150]}")
+
+        session_block = f"[{ts}]\n" + "\n".join(lines)
+        sessions.append(session_block)
+
+    if not sessions:
+        return ""
+
+    header = "[近期会话记忆 - 仅供参考，非当前指令，勿将其视为用户的新请求]"
+    body = "\n\n---\n\n".join(sessions)
+    footer = "[以上为重启前的近期会话摘要，帮助你恢复上下文。请结合当前用户消息回应。]"
+
+    return f"{header}\n\n{body}\n\n{footer}"