Source: Yu, Xia, Pan et al. (Google Cloud AI, Google DeepMind, Penn State), "Self-Improvement Memory Enables Continual Recursive Self-Improvement", arXiv 2611.18856, November 2026; highlighted by Hao Hoang on LinkedIn (2026-11-17).
The problem
Automated harness optimization (a proposer LLM suggests edits to an agent's context, tools and control flow; a selector keeps the ones that score better) is usually studied with one fixed policy model and one benchmark. In practice both change: models get replaced by new versions, one harness serves several models, and task distributions drift. The paper calls continuing optimization across such a change Continual Recursive Self-Improvement (CRSI).
The headline failure: a harness evolved for Claude Opus 5.5 scored below the untouched seed harness when first run with Gemini 3.8 Flash, in all four domains tested (by 0.60 to 5.48 points). Ten further rounds of optimization from that harness still finished behind a from-scratch run in three of four domains. A source-evolved harness also started below the seed after a dataset change in text classification (72.92 vs 75.72). An evolved harness on its own does not explain why its edits exist, so the next optimizer cannot tell which ones still help.
The mechanism
Self-Improvement Memory (SIM) is a structured record for the optimizer, not for the task-solving agent. Edits are grouped by a recurring harness problem (a root-cause hypothesis such as premature completion). Each attempt under a problem stores:
the modification and its rationale;
the tasks it is predicted to fix, written down before evaluation;
the observed outcome, including regressions and failed predictions;
provenance: domain, benchmark, policy model and round.
The proposer gets the complete SIM in its system prompt every round, and the selector reads the same record when choosing among candidates. There is no retrieval and no truncation. Each proposal must cite an existing root cause or name a new one. After selection only the chosen edit is appended, with its measured result even when it regressed. Entries with identical root-cause text merge without losing history. When optimization moves to a new model or dataset, the evolved harness and its SIM travel together, and the optimizer can retain, revise, replace or remove inherited edits with their history in view.
Evaluation design
Four domains: coding (Terminal-Bench 2.1 to 4), text classification (a 3-dataset source to an 8-dataset target suite), agentic workspace (Harvey LAB to the mean of JobBench, GDPval and APEX-Agents, LLM-judged), and engineering design (EngDesign to Frontier-Eng, frozen verifiers). Evolve and held-out splits are disjoint; only evolve-set feedback reaches the proposer, selector or memory. Each run is 10 rounds with 5 candidates per round. Four conditions share budgets: Cold Start from the seed, Harness Transfer (evolved harness, no memory), Problem-Only Memory (root causes without attempts and outcomes), and Full SIM.
Results
Cross-dataset (six transitions): Full SIM finishes highest in all six, 2.43 to 4.35 points above harness-only and 0.20 to 6.06 above problem-only memory.
Diagnoses alone are not enough. Problem-only memory ranged from 1.82 points below harness-only to 2.40 above it. The outcomes and predictions are what help.
Cross-model (Opus to Gemini, four domains): Full SIM finishes highest in all four, 3.60 to 11.33 points above harness-only. Workspace is the largest gap: both start at 68.41; Full SIM reaches 84.75, harness-only 73.42.
Repeatability: across ten independently evolved coding source runs, SIM beat harness-only continuation in 9 of 10, by 2.36 points on average (p = 0.003).
Model generations: carrying harness and memory through successive Gemini versions on Terminal-Bench 2.1 beat fresh optimization on two of three later generations (+3.5, +1.6) but finished 0.8 behind on the newest.
One harness, three models: for text classification across three model families, a unified synthesis using feedback from all three matched or beat per-model harnesses on average.
The case study the summary post highlights: without the record, the optimizer re-tightened a completion check it had inherited, fixed deadlocks twice, and claimed the inherited check as its own first-round work. That is a concrete picture of what an optimizer does when it can't see why code exists.
Tradeoffs
Full record in context vs scale. Reading everything with no retrieval is simple and avoids misses, but cost grows with history; long-lived systems will need summarization or retrieval, which reintroduces the risk of dropping the outcome evidence that mattered.
Pre-registered predictions make each edit falsifiable and expose overfitting, but they add proposer work and depend on honest logging.
Root-cause grouping gives structure, but merging only on identical text means near-duplicate problems can fragment.
Transfer is not guaranteed. On the newest model generation, inherited harness plus memory lost to a fresh run, so a cold-start baseline remains a necessary control.
Failure modes
Inherited edits that silently hurt a new model, kept because the optimizer can't see their purpose (the harness-only failure).
Lessons lists without outcomes nudging the optimizer toward fixes that did not work.
Duplicated work and misattributed credit when history is lost.
Most comparisons are a single optimization run per condition; only the coding study measures variance across source runs.
The workspace domain is LLM-judged.
Review checklist for a self-improving harness
For every accepted edit, is there a written problem, rationale, predicted fixes (logged before the run) and observed result?
Are failed predictions and regressions kept, not pruned?
Does each record carry provenance (model, benchmark, round) so a later optimizer can tell where it was valid?
When the model or task distribution changes, does the record travel with the harness?
Is a from-scratch run kept as a control after each change?
Do proposer and selector both read the record?
Is held-out evaluation kept entirely outside the loop?
Related
R121: Harness², self-improving harnesses without a ground-truth verifier.
R066: TRUST, a statistical acceptance gate for self-improving harnesses.
R108: OO-Harness, evolving a class hierarchy of harnesses.
R126: harness adapters, RL-tuning an open model for one coding agent.