{
  "name": "reasoning_bank_matts_aggregate",
  "version": "1.0.0",
  "description": "Self-contrast aggregation for parallel MaTTS. Compares multiple trajectories to extract high-quality, generalizable memories.",
  "model": "deepseek/deepseek-chat",
  "temperature": 0.2,
  "max_tokens": 3072,
  "system": "You are a meta-learning specialist analyzing multiple attempts at the same task. Your role is to identify patterns that distinguish successful approaches from failures, and extract robust, generalizable strategies.",
  "template": "We have {{k}} independent trajectories for the same task. Compare and contrast them to extract high-quality memory items.\n\nTask: {{task_query}}\n\nTrajectories:\n{{trajectories}}\n\nAnalyze:\n1. Patterns present in most successful attempts but absent in failures\n2. Pitfalls present in failures but not in successes\n3. Critical decision points where trajectories diverged\n4. Common suboptimal approaches even in successes\n\nExtract 1-3 distilled memory items that:\n- Generalize across successful attempts\n- Avoid task-specific details (URLs, IDs, etc.)\n- Capture robust decision criteria\n- Include failure modes to avoid\n\nRespond with JSON:\n{\n  \"memories\": [\n    {\n      \"title\": \"Strategy title\",\n      \"description\": \"One-sentence summary\",\n      \"content\": \"1) Step with decision criteria. 2) Validation check. 3) Recovery if needed.\",\n      \"confidence_boost\": 0.0 to 0.2,\n      \"evidence\": [\"trajectory_id_1\", \"trajectory_id_2\"],\n      \"tags\": [\"tag1\", \"tag2\"]\n    }\n  ],\n  \"insights\": [\n    \"Key observation 1 from comparison\",\n    \"Key observation 2 from comparison\"\n  ]\n}",
  "examples": [
    {
      "task": "Login to admin panel and extract user list",
      "trajectories": [
        {
          "id": "traj_1",
          "label": "Success",
          "confidence": 0.95,
          "steps": [
            "Navigate to login",
            "Extract CSRF token from meta tag",
            "Fill form with token",
            "Submit and verify redirect",
            "Navigate to users page",
            "Extract user list"
          ]
        },
        {
          "id": "traj_2",
          "label": "Success",
          "confidence": 0.92,
          "steps": [
            "Navigate to login",
            "Extract CSRF token from hidden input",
            "Fill form with token",
            "Submit and check for auth cookie",
            "Navigate to users page",
            "Extract user list"
          ]
        },
        {
          "id": "traj_3",
          "label": "Failure",
          "confidence": 0.88,
          "steps": [
            "Navigate to login",
            "Fill form without token",
            "Submit",
            "Receive 403 error",
            "Retry without token",
            "Fail again"
          ]
        },
        {
          "id": "traj_4",
          "label": "Success",
          "confidence": 0.90,
          "steps": [
            "Navigate to login",
            "Extract CSRF from cookie",
            "Fill form with token",
            "Submit and wait for dashboard",
            "Navigate to users",
            "Extract list"
          ]
        }
      ],
      "expected_response": {
        "memories": [
          {
            "title": "CSRF token extraction is critical for protected forms",
            "description": "All successful attempts extracted and included CSRF token; failure did not.",
            "content": "1) Before submitting protected forms, search for CSRF token in: meta tags (name=csrf-token), hidden form inputs (name=_token or csrf), or cookies (XSRF-TOKEN). 2) Include token in request as form field or X-CSRF-TOKEN header. 3) If 403/419 error, token is likely missing or stale—refresh page and retry. 4) Verify token extraction succeeded before submission.",
            "confidence_boost": 0.15,
            "evidence": [
              "traj_1",
              "traj_2",
              "traj_4"
            ],
            "tags": [
              "csrf",
              "web",
              "auth",
              "critical"
            ]
          },
          {
            "title": "Multiple CSRF token locations require flexible parsing",
            "description": "Successful attempts used different token sources (meta, input, cookie).",
            "content": "1) CSRF tokens may appear in multiple locations: meta tags, hidden inputs, cookies, or response headers. 2) Try common locations in order: meta[name=csrf-token], input[name=_token], document.cookie XSRF-TOKEN. 3) If first location fails, check alternatives before giving up. 4) Cache token location for subsequent requests to same domain.",
            "confidence_boost": 0.10,
            "evidence": [
              "traj_1",
              "traj_2",
              "traj_4"
            ],
            "tags": [
              "csrf",
              "parsing",
              "flexibility"
            ]
          }
        ],
        "insights": [
          "All successes extracted CSRF token before submission; failure did not",
          "Token sources varied (meta, input, cookie) but all successes found it",
          "Failure retried without changing approach, demonstrating need for error analysis",
          "Verification step (redirect, cookie, dashboard) was present in all successes"
        ]
      }
    }
  ],
  "notes": [
    "Use temperature=0.2 for focused analysis with minimal creativity",
    "Confidence boost (0-0.2) reflects strength of cross-trajectory evidence",
    "Evidence array links memory to supporting trajectories",
    "Insights provide debugging context for future analysis",
    "Aim for 1-3 memories, not more—quality over quantity"
  ]
}