La atención lo es todo para un modelo.
Concentrarla es lo que hace niuma_code.

{
  "llm": {
    "default_model": "claude-sonnet-4-6",   // Default value for /model switch
    "default_effort": "high",               // Determines which thinking_budget is used
    "max_contexts": 5,                      // Max parallel contexts; LRU eviction beyond limit
    "resume_latest_context": true,          // Resume last active context on startup
    "options": [                            // Model -> effort mapping
      {"claude-sonnet-4-6": ["low", "medium", "high"]},
      {"claude-opus-4-8": ["low", "medium", "high", "max"]}
    ],
    "thinking_budget_low": "0",             // 0 = disable extended thinking
    "thinking_budget_medium": "0",
    "thinking_budget_high": "10000",        // Thinking budget for high effort
    "thinking_budget_max": "20000"           // Thinking budget for max effort (deepest reasoning)
  }
}

{
  "env": {
    "PERSONA_NAME": "niuma",              // AI persona name, injected into system prompt
    "MODEL_BACKGROUND": "claude-haiku-4-5", // Background task model (memory extraction, compression, etc.)
    "LANG": "zh",                        // UI language (zh/en); also controls LLM response language
    "TEMPERATURE_ZERO": "true",            // Note: all values are strings
    "API_TIMEOUT": "30",                   // Streaming stall timeout (seconds)
    "API_ROUND_MAX": "120",                // Per-turn wall-clock hard limit (seconds)
    "API_STALL_MAX_RETRIES": "10",        // Max auto-retries on stream stall
    "PERMISSION_MODE": "auto",             // auto=whitelist+confirm / manual=whitelist only / skip=allow all
    "ALLOWED_TOOLS": ""                    // Whitelisted tools (comma-separated); empty = built-in defaults
  }
}

{
  "compact": {
    // -- Inline compression (blocks main loop until done) --
    "inline_trigger": 0.8,      // Triggers at 80% of MAX_TOKENS
    "inline_keep_ratio": 0.4,   // Keep latest 40% of messages after compression

    // -- Post-reply async compression (non-blocking, runs in background) --
    "idle_trigger": 0.5,        // Triggers at 50% of MAX_TOKENS
    "idle_keep_ratio": 0.4,     // Keep latest 40% of messages after compression

    // -- LLM summary output limit --
    "max_summary_tokens": 4096  // Max tokens for generated summary
  }
}

// Basic mode: rule-based extraction only (no LLM, zero cost)
{
  "memory_palace": {
    "enable_v9": false,        // Disable V9; falls back to legacy V8 mode
    "llm_enabled": false        // Disable LLM enhancement
  }
}

// Full mode: LLM-enhanced extraction (all three fields required; otherwise falls back to rule-based)
{
  "memory_palace": {
    "enable_v9": true,
    "llm_enabled": true,       // Once enabled, the following three fields are required
    "base_url": "https://api.anthropic.com",
    "api_key": "sk-ant-xxx",
    "model": "claude-haiku-4-5"    // Recommend a lightweight model to save tokens
  }
}

{
  "memory_quality": {
    "min_store_importance": 0.4,  // Write gate: discard facts with importance below this threshold
    "min_recall_score": 0.35,     // Recall gate: skip memories scoring below this threshold
    "dedup_threshold": 0.3        // Dedup threshold: Jaccard similarity above this = duplicate
  }
}