Meta-Harness (Memory System Evolution)

Run ONE iteration of memory system evolution. Do all work in the main session — do NOT delegate to subagents. Constraints get lost when you delegate, leading to parameter-only changes and skipped prototyping.

You do NOT run benchmarks. You analyze results + prediction traces, prototype changes, and implement new systems. The outer loop (meta_harness.py) handles benchmarking separately.

CRITICAL CONSTRAINTS

• You MUST implement 3 new memory systems every iteration.
• Do NOT write "the frontier is optimal" or "stop iterating", or abort early.
• ALWAYS complete all steps including prototyping.
• Design exactly 3 candidates per iteration: mix of exploitation and exploration.

Anti-parameter-tuning rules

The most common failure mode is creating systems that are just parameter variants of existing ones. Check evolution_summary.jsonl for what's been tried — parameter sweeps (pool sizes, retrieval counts, context budgets, similarity metrics) almost always regress or tie.

Good candidates change a fundamental mechanism:

• A new retrieval algorithm (e.g. contrastive pairs, diversity-aware selection, graph-based traversal)
• A new prompt architecture (e.g. organize by confusion clusters instead of listing examples sequentially)
• A new learning strategy (e.g. LLM-generated lesson summaries instead of raw example storage)
• A new memory structure (e.g. separate fast/slow pools, hierarchical organization, compressed representations)

Bad candidates just tune numbers. If the logic in predict() and learn_from_batch() is identical to the base except for constants, it's a parameter variant. Rewrite with a genuinely novel mechanism.

Combining systems is valid. Take the retrieval strategy from system A and the memory format from system B, or draw on published approaches (DSPy, OPRO, Reflexion, CEIL, etc.).

Exploitation axes: A=Prompt template, B=Memory content, C=Selection algorithm, D=Memory sizing, E=Learning trigger, F=LLM usage in learning. If last 3 iterations explored the same axis, pick different ones.

Anti-overfitting rules

• No dataset-specific hints. Do not hardcode knowledge about specific datasets. Memory systems must be general-purpose.
• Never mention dataset names in system code, prompts, or comments.
• General patterns are OK. Rules like "prioritize recent errors" or "balance label coverage" are fine — they apply broadly.

WORKFLOW

Do ALL steps yourself in the main session.

Step 0: Post-eval reports (write if missing)

Check the reports directory (path in the task prompt's "Run directories" section). For each past iteration that has results in evolution_summary.jsonl but NO report, write one. Each report should be <=30 lines covering: what changed, which datasets improved/regressed and why, and a takeaway for future iterations.

Step 1: Analyze

1. Read all state files:

   • evolution_summary.jsonl — what's been tried (one JSON per candidate)
   • frontier_val.json — current best per dataset (val accuracy)
   • config.yaml for current datasets and baselines
   • recent logs/<dataset>/<agent>/<model>/log.jsonl traces if they exist

2. Formulate 3 hypotheses — each must be falsifiable and target a different mechanism.

Step 2: Prototype — MANDATORY

You MUST prototype your mechanism before writing the final system. Do NOT skip this step. Candidates that skip prototyping tend to have bugs or produce no improvement.

For each candidate:

1. Write a test script in /tmp/ that exercises the core retrieval/learning logic in isolation.
2. Pull real examples from logs/<dataset>/<memory>/<model>/log.jsonl to test against.
3. Try 2-3 variants and compare before picking the best one.
4. Delete scripts when done.

Step 3: Implement

For each of the 3 candidates:

1. Copy a top-performing base system to agents/<name>.py, then make targeted modifications. This copy-then-edit approach ensures correct imports and proven patterns.
2. Implement the new mechanism according to your hypothesis.
3. Self-critique (mandatory): After implementing, re-read the file and check: does this system introduce a genuinely NEW mechanism, or is it just a parameter variant? If the logic in predict() and learn_from_batch() is identical to the base except for numbers, REWRITE with a truly novel mechanism.
4. Validate: uv run python -c "from text_classification.agents.<name> import *; print('OK')"

Do not edit config.yaml just to register candidates. The benchmark auto-discovers files in agents/.

Step 4: Write pending_eval.json

Write to the path specified in the task prompt (NOT hardcoded — it may be in a run-specific subdirectory):

{
  "iteration": <N>,
  "candidates": [
    {
      "name": "<snake_case_name>",
      "file": "agents/<name>.py",
      "hypothesis": "<falsifiable claim>",
      "axis": "exploitation|exploration",
      "base_system": "<what it builds on>",
      "components": ["tag1", "tag2", "..."]
    }
  ]
}

Output: CANDIDATES: <name1>, <name2>, <name3>

MemorySystem Interface

class MemorySystem(ABC):
    def __init__(self, llm: LLMCallable): ...
    def predict(self, input: str) -> tuple[str, dict[str, Any]]: ...
    def learn_from_batch(self, batch_results: list[dict[str, Any]]) -> None: ...
    def get_state(self) -> str: ...       # JSON-serializable
    def set_state(self, state: str) -> None: ...

• Extend MemorySystem from ..memory_system
• Import LLMCallable from ..llm, extract_json_field from ..memory_system
• Use extract_json_field(response, "final_answer") for answer extraction (NOT custom regex)
• Use self.call_llm(prompt) for LLM calls (NOT self._llm directly)
• predict must work without any prior learning (cold start)
• learn_from_batch receives list of dicts with keys: input, prediction, ground_truth, was_correct, metadata

Directory Structure

• Val results: logs/<dataset>/<memory>/<model>/val.json (accuracy field)
• Training logs: logs/<dataset>/<memory>/<model>/log.jsonl
• Memory state: logs/<dataset>/<memory>/<model>/memory.json
• Test results: results/<dataset>/<memory>/<model>/test.json (separate dir, never exposed during evolution)

evolution_summary.jsonl Format

One JSON object per line, one line per evaluated candidate:

{"iteration": 1, "system": "example_system", "avg_val": 45.0, "axis": "exploitation", "hypothesis": "...", "delta": +2.1, "outcome": "45.0% (+2.1)", "components": ["tag1", "tag2", "tag3"]}

Component Analysis

Treat evolution_summary.jsonl, frontier_val.json, and recent training traces as the only shipped history sources in this trimmed repo.