Meta-Harness as Rework Input

Target: Designing a Memory System for LLM-Based Agents

This is a rework note, not a fresh critique. Meta-Harness does not overturn the current design argument. It hardens a few parts of it and suggests where the next revision should get more operational.

What Meta-Harness strengthens

  • Raw traces are load-bearing. The design note already links the paper ingest for this, but the repo review strengthens the claim with code-grounded evidence: rich traces are not just provenance material; they are the substrate that later improvement loops actually inspect.
  • Oracle quality is the bottleneck. Meta-Harness works because the mutation surface and evaluation loop are explicit. That supports the note's existing claim that corrections and procedure-like patterns are easier to learn from than discoveries.
  • Some durable learning targets should be executable. The design note already allows scripts/tests/checks as promotion targets. Meta-Harness is a clean external example where the learned artifact is harness code, not prose memory.

What it does not settle

  • It does not justify the full four-layer architecture by itself.
  • It does not solve the open-ended discovery oracle problem.
  • It does not show that a general agent memory system should be benchmark-optimized in the same way. It is a workshop optimizer under strong task-specific oracles.

Rework implications

  1. Add a domain-spec section. Before any automated memory-learning loop, the note should name the need for a written domain spec: evaluation unit, fixed vs mutable parts, budget, held-out split, leakage risks, and candidate interface.
  2. Sharpen workshop vs library language. Meta-Harness is best read as a workshop optimizer whose outputs may later be distilled into library artifacts. The note should make that boundary clearer.
  3. State executable promotion more explicitly. When the learned lesson is really about retrieval, context assembly, or action scaffolding, the right target may be code or a check, not a prose memory.
  4. Strengthen the anti-summary point. The note should say more directly that compressed artifacts are not drop-in substitutes for raw traces during debugging, redistillation, or outer-loop improvement.

Working judgment

Meta-Harness should inform the rework as an implementation-level hardening of the current argument, not as a replacement architecture. It strengthens the note's oracle story, trace-retention story, and system-definition-to-code story. The main addition it pushes is: before discussing automated learning from memory, specify the domain and oracle contract of the loop.