LangGraph ReAct agent context window exploding despite ContextEditingMiddleware - need help

## Summary
`ClearToolUsesEdit` middleware does not trim context as expected - agent context keeps growing despite configuration, leading to throttling and performance issues.

## Details

I'm running a LangGraph ReAct agent with multiple tool calls per user query (Knowledge Base searches + SQL executions). Despite configuring `ClearToolUsesEdit`, the context continues to grow to 150K-200K tokens.

**Environment:**
- `langgraph` version: 0.3.x (latest)
- `langchain-aws` version: latest
- Python 3.11
- Running on AWS Bedrock AgentCore Runtime (serverless)
- Model: Claude Haiku (200K context limit)
- Using Bedrock AgenCore Memory (Short term)

**Current Configuration:**

from langgraph.context_editing import ContextEditingMiddleware, ClearToolUsesEdit

context_editor = ContextEditingMiddleware(
edits=[ClearToolUsesEdit(
trigger=100000, # Trigger at 100K tokens
clear_at_least=20000, # Reclaim at least 20K tokens
keep=5, # Keep 5 most recent tool results
clear_tool_inputs=True,
)]
)

agent = create_react_agent(
model=llm,
tools=tools,
prompt=system_prompt,
context_editing=context_editor,
)

**What I've Tried:**
1. Lowering `trigger` from 100K to 50K to 10K
2. Increasing `clear_at_least` to 20K, 30K
3. Reducing `keep` from 5 to 3 to 1
4. Setting `clear_tool_inputs=True`
5. Various combinations of the above

**Observed Behavior:**
- Context grows linearly with each tool call
- By the 5th-6th tool call, context reaches 100K+ tokens
- Agent gets throttled by Bedrock TPM limits
- Trimming doesn't seem to occur, or occurs too late

Questions:

When does trimming actually happen - before or after LLM call?

Does trigger mean "trim when context exceeds this" or something else?

Better alternatives for aggressive context management?