Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

fix(session): honor output language for working memory (#4514)

Yohanes7ab1f965668e
diff --git a/openviking/prompts/templates/compression/ov_wm_v2.yaml b/openviking/prompts/templates/compression/ov_wm_v2.yaml
index cf13c0729..e4b1f945f 100644
--- a/openviking/prompts/templates/compression/ov_wm_v2.yaml
+++ b/openviking/prompts/templates/compression/ov_wm_v2.yaml
@@ -11,6 +11,10 @@ variables:
type: "string"
description: "Session content to summarize into the new Working Memory"
required: true
+ - name: "output_language"
+ type: "string"
+ description: "Language for generated Working Memory section content"
+ required: true
- name: "latest_archive_overview"
type: "string"
description: "Optional prior Working Memory. When provided, still-valid info should be carried forward."
@@ -47,6 +51,7 @@ template: |
Output a complete Working Memory using the 7-section structure below.
RULES:
+ - Write all section content in {{ output_language }}. Keep the 7 canonical English section headers and italic descriptions unchanged so existing Working Memory archives remain machine-detectable.
- All 7 section headers and italic descriptions must appear unchanged.
- Write INFO-DENSE content: names, numbers, dates, locations, exact values.
- For temporal facts, record the EVENT date, not the conversation date.
diff --git a/openviking/prompts/templates/compression/ov_wm_v2_update.yaml b/openviking/prompts/templates/compression/ov_wm_v2_update.yaml
index d6679b5f9..13802f06e 100644
--- a/openviking/prompts/templates/compression/ov_wm_v2_update.yaml
+++ b/openviking/prompts/templates/compression/ov_wm_v2_update.yaml
@@ -15,6 +15,10 @@ variables:
type: "string"
description: "New conversation content since the last update."
required: true
+ - name: "output_language"
+ type: "string"
+ description: "Language for generated Working Memory section content"
+ required: true
- name: "wm_section_reminders"
type: "string"
description: "Dynamic section-size warnings generated from the current Working Memory."
@@ -51,6 +55,8 @@ template: |
# RULE 0 — Produce a continuation-grade Working Memory.
+ Write all generated section content in {{ output_language }}. Keep the 7 canonical English section headers unchanged because the update parser relies on them.
+
The goal is a document that lets a future assistant seamlessly continue the
conversation. Concrete facts (names, dates, decisions, relationships, exact
values) MUST survive. But repetitive events of the same type should be
diff --git a/openviking/session/session.py b/openviking/session/session.py
index e6eed9ecd..2ca9f658c 100644
--- a/openviking/session/session.py
+++ b/openviking/session/session.py
@@ -25,6 +25,7 @@ from openviking.server.identity import RequestContext, Role
from openviking.session.auto_commit_policy import AutoCommitPolicy
from openviking.session.memory.constants import AGENT_EVOLUTION_MEMORY_TYPES
from openviking.session.memory_policy import MemoryPolicy
+from openviking.session.memory.utils.language import resolve_output_language_from_conversation
from openviking.session.retention import (
RETENTION_MODE_TURN_BUDGET,
RetentionPlan,
@@ -4230,6 +4231,17 @@ class Session:
return ""
formatted = self._format_messages_for_wm(messages, checkpoint_requests)
+ language_conversation = "\n".join(
+ f"[{message.role}]: {line}"
+ for message in messages
+ for part in message.parts
+ if isinstance(part, TextPart)
+ for line in part.text.splitlines()
+ if line.strip()
+ )
+ output_language = resolve_output_language_from_conversation(
+ language_conversation, config=get_openviking_config()
+ )
checkpoint_instructions = self._checkpoint_prompt_instructions(len(checkpoint_requests))
vlm = get_openviking_config().vlm
@@ -4270,6 +4282,7 @@ class Session:
"messages": formatted,
"latest_archive_overview": latest_archive_overview or "",
"checkpoint_instructions": checkpoint_instructions,
+ "output_language": output_language,
},
)
if checkpoint_requests:
@@ -4328,6 +4341,7 @@ class Session:
"latest_archive_overview": latest_archive_overview,
"wm_section_reminders": reminders,
"checkpoint_instructions": checkpoint_instructions,
+ "output_language": output_language,
},
)
resp = await vlm.get_completion_async(
@@ -4346,7 +4360,7 @@ class Session:
raise
logger.warning("WM update tool_call failed (%s); falling back to creation prompt", e)
return await self._fallback_generate_wm_creation(
- formatted, messages, latest_archive_overview
+ formatted, messages, latest_archive_overview, output_language
)
has_tc = bool(getattr(resp, "has_tool_calls", False) and getattr(resp, "tool_calls", None))
@@ -4363,7 +4377,7 @@ class Session:
raise ValueError("Working Memory update returned no tool call for checkpoints")
logger.warning("WM update: LLM returned no tool_call; falling back to creation prompt")
return await self._fallback_generate_wm_creation(
- formatted, messages, latest_archive_overview
+ formatted, messages, latest_archive_overview, output_language
)
checkpoint_summaries: tuple[str, ...] = ()
@@ -4464,7 +4478,7 @@ class Session:
e,
)
return await self._fallback_generate_wm_creation(
- formatted, messages, latest_archive_overview
+ formatted, messages, latest_archive_overview, output_language
)
_wm_debug(
@@ -4484,6 +4498,7 @@ class Session:
formatted_messages: str,
messages: List[Message],
prior_overview: str = "",
+ output_language: str = "en",
) -> str:
"""Re-run WM creation prompt when the update tool_call path fails.
@@ -4503,6 +4518,7 @@ class Session:
"messages": formatted_messages,
"latest_archive_overview": prior_overview,
"checkpoint_instructions": "",
+ "output_language": output_language,
},
)
return await get_openviking_config().vlm.get_completion_async(prompt)
diff --git a/tests/session/test_session_retention_integration.py b/tests/session/test_session_retention_integration.py
index 89f13783e..9c8716d9e 100644
--- a/tests/session/test_session_retention_integration.py
+++ b/tests/session/test_session_retention_integration.py
@@ -751,6 +751,70 @@ async def test_wm_creation_returns_two_products_in_one_model_call(
assert "secret-source-id" not in calls[0]["prompt"]
+async def test_wm_creation_passes_configured_output_language_to_prompt(client, monkeypatch):
+ session = client(session_id="wm_output_language_prompt_test")
+ prompts: list[dict] = []
+
+ class FakeVLM:
+ @staticmethod
+ def is_available() -> bool:
+ return True
+
+ async def get_completion_async(self, **kwargs):
+ prompts.append(kwargs)
+ return "# Working Memory"
+
+ vlm = FakeVLM()
+ monkeypatch.setattr(
+ "openviking.session.session.get_openviking_config",
+ lambda: SimpleNamespace(vlm=vlm, output_language_override="zh-CN"),
+ )
+
+ result = await session._generate_archive_summary_async(
+ [_text_message("zh-user", "user", "请总结当前部署状态")]
+ )
+
+ assert result == "# Working Memory"
+ assert len(prompts) == 1
+ assert "zh-CN" in prompts[0]["prompt"]
+
+
+async def test_wm_creation_detects_language_from_multiline_user_message(client, monkeypatch):
+ session = client(session_id="wm_multiline_output_language_detection_test")
+ prompts: list[dict] = []
+
+ class FakeVLM:
+ @staticmethod
+ def is_available() -> bool:
+ return True
+
+ async def get_completion_async(self, **kwargs):
+ prompts.append(kwargs)
+ return "# Working Memory"
+
+ vlm = FakeVLM()
+ monkeypatch.setattr(
+ "openviking.session.session.get_openviking_config",
+ lambda: SimpleNamespace(vlm=vlm),
+ )
+
+ result = await session._generate_archive_summary_async(
+ [
+ _text_message(
+ "multiline-user",
+ "user",
+ "Task details:\n"
+ "当前生产环境已经完成部署。\n"
+ "请用中文总结当前状态和后续风险。",
+ )
+ ]
+ )
+
+ assert result == "# Working Memory"
+ assert len(prompts) == 1
+ assert "zh-CN" in prompts[0]["prompt"]
+
+
async def test_wm_update_returns_two_products_in_one_model_call(
client,
monkeypatch,