fix(memory): bound extraction conversation prompts

docs/en/guides/01-configuration.md

@@ -1364,6 +1364,7 @@ For memory-related settings, add a `memory` section in `ov.conf`:
 | `version` | Deprecated and ignored. OpenViking always uses the v3 memory extraction pipeline; existing configs that set this field still load without error. | `"v3"` |
 | `custom_templates_dir` | Custom memory templates directory. If set, templates from this directory are loaded in addition to built-in templates. | `""` |
 | `extraction_enabled` | Whether session commit runs long-term memory extraction. | `true` |
+| `extraction_prompt_max_chars` | Maximum characters in the conversation-history user message sent to one memory-extraction VLM request. Older messages are omitted first while stable range indices are preserved. System/schema and follow-up tool messages are outside this budget. | `60000` |
 | `session_skill_extraction_enabled` | Whether session commit also extracts reusable skills into the current user's skill directory. | `false` |
 | `link_enabled` | Whether memory extraction writes and resolves memory links. | `false` |
 

docs/zh/guides/01-configuration.md

@@ -1331,6 +1331,7 @@ openviking-server --config /path/to/ov.conf
 | `version` | 已废弃且会被忽略。OpenViking 始终使用 v3 记忆抽取链路；已有配置中保留该字段仍可正常加载，不会报错。 | `"v3"` |
 | `custom_templates_dir` | 自定义 memory templates 目录。设置后会在内置模板之外加载该目录中的模板。 | `""` |
 | `extraction_enabled` | session commit 时是否执行长期记忆抽取。 | `true` |
+| `extraction_prompt_max_chars` | 单次记忆抽取 VLM 请求中，conversation-history user message 的最大字符数。超限时优先省略较早消息，并保留原始 range 索引；system/schema 与后续工具消息不计入该预算。 | `60000` |
 | `session_skill_extraction_enabled` | session commit 时是否同时抽取可复用 skill 到当前用户的 skill 目录。 | `false` |
 | `link_enabled` | 记忆抽取是否写入和解析 memory links。 | `false` |
 

examples/ov.conf.example

@@ -160,6 +160,7 @@
     "memory": {
         "version": "v2",
         "extraction_enabled": true,
+        "extraction_prompt_max_chars": 60000,
         "session_skill_extraction_enabled": false,
     },
     "ingest": {

openviking/session/memory/session_extract_context_provider.py

@@ -48,6 +48,10 @@
 _PREFETCH_SEARCH_TEXT_PART_MAX_CHARS = 1000
 _PREFETCH_SEARCH_ASSISTANT_TEXT_PART_MAX_CHARS = 500
 _PREFETCH_SEARCH_TOOL_FIELD_MAX_CHARS = 500
+_DEFAULT_EXTRACTION_PROMPT_MAX_CHARS = 60000
+_EXTRACTION_PROMPT_OMISSION_MARKER = (
+    "...[conversation content omitted to fit memory.extraction_prompt_max_chars]..."
+)
 _RESOURCE_REASON_LANGUAGE_RE = re.compile(
     r"(?im)^\s*(?:User reason|用户说明|用户原因|用户理由)[:：]\s*(.+?)\s*$"
 )
@@ -80,6 +84,11 @@ def __init__(
         config = get_openviking_config()
         self._eager_prefetch = config.memory.eager_prefetch if config.memory else False
         self._prefetch_search_topn = config.memory.prefetch_search_topn if config.memory else 5
+        self._extraction_prompt_max_chars = (
+            config.memory.extraction_prompt_max_chars
+            if config.memory
+            else _DEFAULT_EXTRACTION_PROMPT_MAX_CHARS
+        )
         self._ctx = ctx
         self._viking_fs = viking_fs
         self._transaction_handle = transaction_handle
@@ -267,21 +276,72 @@ def _build_conversation_message(self) -> Dict[str, Any]:
         else:
             time_display = session_time_str
 
-        extract_context = self.get_extract_context()
-        conversation = self._assemble_conversation(extract_context.messages)
-
-        return {
-            "role": "user",
-            "content": f"""## Conversation History
+        prefix = f"""## Conversation History
 **Session Time:** {time_display} ({day_of_week})
 Relative times (e.g., 'last week', 'next month') are based on Session Time, not today.
 
-{conversation}
+"""
+        suffix = """
+
+After exploring, analyze the conversation and output ALL memory write/edit/delete operations in a single response. Do not output operations one at a time - gather all changes first, then return them together."""
+        extract_context = self.get_extract_context()
+        conversation_budget = max(
+            1,
+            self._extraction_prompt_max_chars - len(prefix) - len(suffix),
+        )
+        conversation = self._assemble_conversation(
+            extract_context.messages,
+            max_chars=conversation_budget,
+        )
+        content = f"{prefix}{conversation}{suffix}"
+        return {"role": "user", "content": content}
 
-After exploring, analyze the conversation and output ALL memory write/edit/delete operations in a single response. Do not output operations one at a time - gather all changes first, then return them together.""",
-        }
+    @staticmethod
+    def _truncate_formatted_message(message: str, max_chars: int) -> str:
+        if len(message) <= max_chars:
+            return message
+        marker = f"\n{_EXTRACTION_PROMPT_OMISSION_MARKER}\n"
+        if max_chars <= len(marker):
+            return message[:max_chars]
+        retained_chars = max_chars - len(marker)
+        head_chars = max(retained_chars // 2, min(len(message), message.find(": ") + 2))
+        head_chars = min(head_chars, retained_chars)
+        tail_chars = retained_chars - head_chars
+        tail = message[-tail_chars:] if tail_chars else ""
+        return f"{message[:head_chars]}{marker}{tail}"
+
+    @classmethod
+    def _bound_formatted_messages(cls, messages: List[str], max_chars: int) -> str:
+        conversation = "\n".join(messages)
+        if len(conversation) <= max_chars:
+            return conversation
+        if not messages:
+            return ""
 
-    def _assemble_conversation(self, messages: Any) -> str:
+        # A single oversized latest message still keeps its stable range header and
+        # both ends of the content. For multi-message histories, prefer complete
+        # recent messages so any emitted range index still maps to the original
+        # ExtractContext entry.
+        if len(messages[-1]) > max_chars:
+            return cls._truncate_formatted_message(messages[-1], max_chars)
+
+        marker = _EXTRACTION_PROMPT_OMISSION_MARKER
+        available = max_chars - len(marker) - 1
+        selected: List[str] = []
+        selected_chars = 0
+        for message in reversed(messages):
+            separator_chars = 1 if selected else 0
+            if selected_chars + separator_chars + len(message) > available:
+                break
+            selected.insert(0, message)
+            selected_chars += separator_chars + len(message)
+
+        if not selected:
+            return cls._truncate_formatted_message(messages[-1], max_chars)
+        selected_text = "\n".join(selected)
+        return f"{marker}\n{selected_text}"
+
+    def _assemble_conversation(self, messages: Any, *, max_chars: Optional[int] = None) -> str:
         """Assemble conversation string from messages.
 
         Args:
@@ -303,8 +363,6 @@ def format_message_with_parts(msg: Message) -> str:
             user utterances and can leak environment/database state into user
             memories. Agent-scope providers enable tool evidence explicitly.
             """
-            from openviking.message.part import ToolPart
-
             parts = getattr(msg, "parts", [])
             formatted_parts: List[str] = []
             for part in parts:
@@ -338,7 +396,18 @@ def format_message_header(msg: Message, idx: int) -> str | None:
             for idx, msg in enumerate(messages)
             if (formatted := format_message_header(msg, idx)) is not None
         ]
-        conversation_sections.append("\n".join(formatted_messages))
+        if max_chars is not None:
+            bounded = self._bound_formatted_messages(formatted_messages, max_chars)
+            if len("\n".join(formatted_messages)) > len(bounded):
+                logger.warning(
+                    "Memory extraction conversation exceeded the configured prompt budget; "
+                    "keeping bounded recent context (%d -> %d chars)",
+                    len("\n".join(formatted_messages)),
+                    len(bounded),
+                )
+            conversation_sections.append(bounded)
+        else:
+            conversation_sections.append("\n".join(formatted_messages))
 
         return "\n\n".join(section for section in conversation_sections if section)
 

openviking_cli/utils/config/memory_config.py

@@ -36,6 +36,16 @@ class MemoryConfig(BaseModel):
             "0 means unlimited retries."
         ),
     )
+    extraction_prompt_max_chars: int = Field(
+        default=60000,
+        ge=1000,
+        description=(
+            "Maximum characters in the conversation-history user message sent for one "
+            "memory extraction request. System/schema and tool follow-up messages are not "
+            "included in this limit. Older messages are omitted first while original range "
+            "indices are preserved."
+        ),
+    )
     experimental_memory_switch: bool = Field(
         default=False,
         description=(

tests/session/memory/test_memory_react_system_prompt.py

@@ -5,7 +5,10 @@
 """
 
 from openviking.message import ImagePart, Message, TextPart, ToolPart
-from openviking.session.memory.session_extract_context_provider import SessionExtractContextProvider
+from openviking.session.memory.session_extract_context_provider import (
+    _EXTRACTION_PROMPT_OMISSION_MARKER,
+    SessionExtractContextProvider,
+)
 from openviking.session.memory.vision_message_normalizer import IMAGE_DESCRIPTION_PROMPT
 
 
@@ -87,6 +90,39 @@ def test_instruction_includes_resource_uri_handling_for_user_scoped_resource_uri
 
 
 class TestSessionConversationToolFiltering:
+    def test_extraction_prompt_omits_older_messages_within_configured_budget(self):
+        messages = [
+            Message(id="m0", role="user", parts=[TextPart("old-" + "a" * 700)]),
+            Message(id="m1", role="assistant", parts=[TextPart("middle-" + "b" * 700)]),
+            Message(id="m2", role="user", parts=[TextPart("latest-" + "c" * 700)]),
+        ]
+        provider = SessionExtractContextProvider(messages=messages)
+        provider._extraction_prompt_max_chars = 1200
+
+        prompt = provider._build_conversation_message()["content"]
+
+        assert len(prompt) <= 1200
+        assert _EXTRACTION_PROMPT_OMISSION_MARKER in prompt
+        assert "[2][user][user]: latest-" in prompt
+        assert "[0][user][user]: old-" not in prompt
+        assert "[1][assistant][assistant]: middle-" not in prompt
+
+    def test_extraction_prompt_truncates_one_oversized_message_and_keeps_range_header(self):
+        message = Message(
+            id="m0",
+            role="user",
+            parts=[TextPart("begin-" + "x" * 4000 + "-end")],
+        )
+        provider = SessionExtractContextProvider(messages=[message])
+        provider._extraction_prompt_max_chars = 1200
+
+        prompt = provider._build_conversation_message()["content"]
+
+        assert len(prompt) <= 1200
+        assert "[0][user][user]: begin-" in prompt
+        assert "-end" in prompt
+        assert _EXTRACTION_PROMPT_OMISSION_MARKER in prompt
+
     def test_session_conversation_omits_skill_tool_call(self):
         messages = [
             Message(
@@ -232,8 +268,6 @@ def test_detect_language_prefers_user_text_over_assistant_text(self):
 
         assert provider._detect_language() == "zh-CN"
 
-
-
     async def test_prepare_extraction_messages_replaces_image_part_with_vlm_description(self):
         class FakeVisionVLM:
             def __init__(self):
@@ -341,6 +375,7 @@ async def test_prepare_extraction_messages_does_not_replace_caller_message_list(
         assert any(isinstance(part, ImagePart) for part in messages[0].parts)
         assert provider.messages is not messages
 
+
 def test_session_provider_empty_messages_still_uses_environment_fallback(monkeypatch):
     monkeypatch.setenv("TZ", "Asia/Shanghai")
     provider = SessionExtractContextProvider(messages=[])