Update context management for MiroThinker v1.5

[jenny-miromind]

MiroThinker v1.5 Context Management Strategy

MiroThinker v1.5 employs a two-layer context management mechanism, each addressing different dimensions of context challenges.

Overview

Layer	Config Parameter	Purpose
Layer 1	keep_tool_result	Controls context growth within a single attempt
Layer 2	context_compress_limit	Compresses context window and enables knowledge transfer across attempts

Layer 1: keep_tool_result - Recency-based Context Retention

Core Problem

Traditional ReAct stuffs all tool outputs into the message history, leading to:

• Context bloat
• Early tool outputs crowding out space
• Low utilization efficiency

Solution

Config Value	Behavior
keep_tool_result = K	Keep only the most recent K tool responses, while preserving the complete thoughts & actions trace
keep_tool_result = -1	Keep all tool results (no trimming)

Benefits

• Focus on recent observations
• Free up context space
• Support longer interaction/tool call trajectories

Layer 2: context_compress_limit - Context Compression

Core Problem

Even with keep_tool_result trimming, a single attempt may still fail to complete the task due to:

• Running out of turns (max_turns)
• Reaching context window limits

Solution

Config Value	Behavior
context_compress_limit = 0	Disable context compression retry; only one chance, fallback to intermediate answers on failure
context_compress_limit = N	Enable context compression retry; up to N retries

Relation to DeepSeek V3.2's "Discard-all" Strategy

This mechanism is conceptually similar to DeepSeek V3.2's "Discard-all" strategy, which resets the context by discarding all previous tool call history. Both approaches share the same core insight:

A fresh start with learned experience is often more effective than continuing with a bloated context.

How It Works

When a task is not completed within the given turns and context window:

Step 1: Generate Failure Experience Summary (context compression)

Compress conversation history into ~500-1000 tokens structured summary containing:

• Failure type: incomplete / blocked / misdirected / format_missed
  • incomplete: ran out of turns before finishing
  • blocked: got stuck due to tool failure or missing information
  • misdirected: went down the wrong path
  • format_missed: found the answer but forgot to use \boxed{}
• What happened: the approach taken and why a final answer was not reached
• Useful findings: facts, intermediate results, or conclusions to be reused

Step 2: Inject Summary on Retry

• Summary is injected into the original user query
• Model starts fresh with a clean context window
• Learns from past failures, avoids repeating mistakes

How the Two Layers Work Together

┌─────────────────────────────────────────────────────────────────────────────┐
│                          Single Task Attempt                                │
│  ┌─────────────────────────────────────────────────────────────────────┐    │
│  │  keep_tool_result = K                                               │    │
│  │  ════════════════════                                               │    │
│  │  • Keep only the most recent K tool responses                       │    │
│  │  • Preserve complete thoughts & actions trace                       │    │
│  │  → Controls context growth within a single attempt                  │    │
│  └─────────────────────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼ Task not completed?
┌─────────────────────────────────────────────────────────────────────────────┐
│                       context_compress_limit > 0                            │
│  ════════════════════════════════════════════════                           │
│  1. Generate Failure Experience Summary (compress context)                  │
│  2. Inject summary into task description                                    │
│  3. Start fresh with clean context                                          │
│  → Enables knowledge transfer across attempts                               │
└─────────────────────────────────────────────────────────────────────────────┘