AgentCore Memory + Browser Tool: Conversation history corruption causing orphaned toolUse

未關閉
#245 1 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
4/5
預估耗時
3-5 天
新手友好度
35/100
Issue 類型
缺陷
描述清晰度
基本清楚
活躍度
停滯
技術堆疊
aws, python

研究方向

從 AgentCoreMemorySessionManager 開始,使用提供的 Browser Tool 和記憶體設定重現故障,並比較儲存前與載入後的對話歷史。第二次 agent 呼叫成功,且載入後的歷史保留有效的 toolUse/toolResult 配對,即表示完成;issue 中未指定特定檔案或測試。

由索引模型根據 Issue 內容生成。

描述

Bug Description

When using AgentCoreMemorySessionManager with AgentCoreBrowser (Browser Tool) in Strands Agents, the conversation history becomes corrupted after the first message, leading to ValidationException errors.

Environment

  • bedrock-agentcore: Latest version
  • Strands Agents: Latest version
  • strands-agents-tools: Latest version
  • Region: us-east-1
  • Memory ID: memory_agentcore_demo_complex-MOvGs2AZQ5

Problem Description

Error Message
Session message history has an orphaned toolUse with no toolResult. 
Adding toolResult content blocks to create valid conversation.

ValidationException: The number of toolResult blocks at messages.27.content 
exceeds the number of toolUse blocks of previous turn.
Behavior
  • ✅ First message in session: Works correctly
  • ❌ Subsequent messages in same session: Fails with ValidationException
  • ✅ Without Browser Tool: Memory works correctly with other tools (Gateway tools)
  • ❌ With Browser Tool: Memory appears to save corrupted conversation history

Code Example

from bedrock_agentcore.memory.integrations.strands.config import AgentCoreMemoryConfig
from bedrock_agentcore.memory.integrations.strands.session_manager import AgentCoreMemorySessionManager
from strands import Agent
from strands_tools.browser import AgentCoreBrowser

# Memory configuration
memory_config = AgentCoreMemoryConfig(
    memory_id="memory_agentcore_demo_complex-MOvGs2AZQ5",
    session_id="test_session_123",
    actor_id="test_user"
)

session_manager = AgentCoreMemorySessionManager(
    agentcore_memory_config=memory_config,
    region_name="us-east-1"
)

# Browser Tool
browser_tool = AgentCoreBrowser(region="us-east-1")

# Agent with Browser Tool + Memory
agent = Agent(
    model="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    system_prompt="You are a helpful assistant.",
    session_manager=session_manager,
    tools=[browser_tool.browser]
)

# First call: SUCCESS ✅
result1 = agent("Hello")

# Second call in same session: FAILS ❌
# Memory loads corrupted history with orphaned toolUse
result2 = agent("How are you?")

Root Cause Hypothesis

The issue appears to be in how AgentCoreMemorySessionManager saves and loads conversation history when Browser Tool is involved:

Possible Causes:
  1. Incomplete tool call persistence:

    • Browser Tool creates tool_use blocks
    • Memory saves the conversation
    • On reload, tool_use blocks are missing or corrupted
    • But tool_result blocks are preserved
    • This creates "orphaned toolUse"
  2. Serialization/Deserialization issue:

    • Browser Tool's conversation format may not be fully compatible with Memory's storage format
    • Some data is lost during save/load cycle
  3. Context truncation without cleanup:

    • Memory may truncate conversation history
    • But doesn't properly clean up orphaned tool blocks

Logs

2026-02-02T10:12:24.909000+00:00 Session message history has an orphaned toolUse with no toolResult. Adding toolResult content blocks to create valid conversation.

2026-02-02T10:12:25.876000+00:00 botocore.errorfactory.ValidationException: An error occurred (ValidationException) when calling the ConverseStream operation: The number of toolResult blocks at messages.27.content exceeds the number of toolUse blocks of previous turn.

Testing Results

Configuration Result
Memory + Gateway Tools ✅ Works
Memory + Browser Tool ❌ Fails after first message
No Memory + Browser Tool ❓ Not tested (Memory required)
Memory + Gateway + Browser ❌ Fails after first message

This strongly suggests the issue is in how Memory handles Browser Tool's conversation history.

Questions

  1. Is Browser Tool + Memory a tested and supported combination?
  2. Are there known limitations or special handling required for Browser Tool with Memory?
  3. Does Memory have special serialization logic for different tool types?
  4. Is there documentation on conversation history format requirements?

Expected Behavior

AgentCore Memory should correctly save and restore conversation history when Browser Tool is used, maintaining proper tool_use/tool_result pairing.

Actual Behavior

After the first message, Memory loads corrupted conversation history with orphaned tool blocks, causing Bedrock API to reject subsequent requests.

Related Issues

  • Similar issue reported in Strands Agents: strands-agents/sdk-python#1610
  • Anthropic Claude Code #13964: orphaned tool_result blocks after context truncation
  • browser-use #710: Browser-use fails with Bedrock Converse API

Suggested Investigation

  1. Add logging to show exactly what Memory saves vs. what it loads
  2. Compare conversation history format for Browser Tool vs. other tools
  3. Check if Browser Tool's message format requires special handling
  4. Verify that all tool blocks are properly paired during save/load

Suggested Fix

Add validation and cleanup in AgentCoreMemorySessionManager:

  1. Before saving: Validate that all tool_use blocks have corresponding tool_result blocks
  2. After loading: Sanitize conversation history to remove orphaned tool blocks
  3. Add specific handling for Browser Tool if its format differs from standard tools

Impact

This bug prevents using Browser Tool with Memory, which is a critical limitation for building context-aware agents that need web browsing capabilities.

主要語言
Python
星號
786
分支
163
平均合併
4 小時
30 天內合併 PR
22

環境準備

  • 沒有 Dockerfile 或 Docker Compose 檔案
  • 沒有 Pull Request 範本
  • 閱讀貢獻指南

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

aws/bedrock-agentcore-sdk-python 的其他 Issue

查看 aws/bedrock-agentcore-sdk-python 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。