Continual recursive self-improvement

Keep the code.
Remember why.

When the model or benchmark changes, carry forward the experience behind your agent’s improvements.

Self-Improvement Memory Enables
Continual Recursive Self-Improvement · Preprint, 2026

EXPERIENCE / CARRIED FORWARD
01 / THE ARTIFACTEvolved harness
+
02 / THE EXPERIENCEWhy we changed itWhat we expectedWhat actually happened
NEW MODEL OR BENCHMARKContinue improvingUpdate harness + memory
POLICY WEIGHTS: FROZENMEMORY: EXTENSIBLE
+2.36Mean score-point gain
in paired coding continuations
9 / 10Pairs favor SIM
One negative pair retained
4Task domains
Cross-dataset & cross-model

Xin Yu1,4 · Fei Xia1 · Jiaqi Pan1 · Youbao Tang1 · Chengzhi Liu1 · Adams Yu2 · Sicong Jiang2 · Yanfei Chen3 · Lingzhou Xue4 · Ting Yu1

¹ Google Cloud AI   ² Google DeepMind   ³ Google Cloud AI Research   ⁴ The Pennsylvania State University

Work done during Xin Yu’s internship at Google.

01 / The continuity problem

Good code.
Missing context.

An agent’s harness organizes its prompts, tools, context and execution. Optimizing that layer produces more than code: it produces reasons, failed attempts and evidence. A model or task change makes that history useful again.

CARRY ONLY THE HARNESS

“Why is this check here?”

if command_succeeded:
  verify_completion()

The code survives. The diagnosis, rejected alternatives and conditions behind the edit may not. The next optimizer has to reconstruct its purpose.

CARRY THE HARNESS + SIM

“Keep the reason. Revise the check.”

Problem → Edit → Prediction
→ Observed outcome → Context

A completion check can retain its rationale while adapting to a new task’s constraints. New evidence extends the record instead of replacing its history.

CRSI continues an existing optimization run after a compatible benchmark or policy-model change. SIM is memory for the harness optimizer, rather than memory supplied to the task-solving policy.

02 / Inside the memory

An edit has a history.
Let the next run read it.

SIM groups selected edits by recurring harness problem. Both the proposer and the selector read the record. Further attempts add to it—even when a prediction misses or an edit regresses.

ONE RECORD / THREE VIEWS

Premature completion

ILLUSTRATIVE RECORD

A compact example based on the paper’s completion-verification case. This is an explanation, not a raw exported SIM entry.

PROBLEM
A successful command does not necessarily mean the requested task is complete.
RATIONALE
Keep completion verification tied to the deliverable, rather than the agent’s declaration.
NEXT QUESTION
Which checks still apply when the target task changes?
01 / READ

Recover the context

Read earlier edits, outcomes and the settings in which they were tested.

02 / PROPOSE

Make the next edit

Refer to an existing problem or identify a new one.

03 / SELECT

Evaluate candidates

Use evolve-set feedback to select a harness revision.

04 / APPEND

Keep the evidence

Record the selected edit and measured outcome; preserve prior attempts.

Held-out evaluation remains outside the optimization loop. The current implementation reads the complete SIM, without retrieval or truncation. Method · §3 ↗

03 / The paired test

Same starting code.
Different histories.

Ten independent Terminus-2 source runs on TB2.1. Each evolved harness initializes two matched, ten-round TB4 continuations: one with SIM and one without. The source run is the unit of analysis.

+2.36 pts

Mean final-score advantage over harness-only continuation. SIM finishes higher in nine pairs; one favors no SIM.

95% CI [1.05, 3.68]
Paired t-test · p = 0.003
10 independent source runs

SIM − NO SIM / SCORE POINTSALL TEN PAIRS

One bar per independent source run. Amber marks the negative pair.
Reported paired mean uses unrounded scores; displayed endpoints are rounded.

Inspect starting and final scores +
Paper Table 3, paired coding continuation scores
RunStartNo SIMSIMΔ
View the original coding result figure ↗Paper Figure 5: coding optimization trajectories and all ten paired score differences, including the negative eighth pair

Source: paper §4.4, Table 3 and Figure 5. Starting harness, target policy, benchmark, rounds and candidate budgets match within each pair. Read the comparison ↗

04 / Across settings

New tasks.
New models.
Continued search.

Explore the direct memory comparison. All three continuation conditions inherit the same source-evolved harness; the task-solving policy, target tasks, rounds and candidate budgets are matched.

These main setting comparisons use one complete optimization run per condition, with five rollouts per task. They describe these settings; they are not ten independent repetitions per domain. Scores use each benchmark’s scale and should not be averaged across domains.

Paper Figures 3–4 & Tables 5–6 ↗
05 / What the evidence supports

Useful memory.
Not a universal guarantee.

The study tests whether a record of earlier optimization helps continued search. It does not establish unbounded self-improvement, universal transfer, or the isolated causal effect of every memory field.

Read the full setup and limitations ↗
  • Compatible interfaces

    The source-evolved harness must remain executable in the target setting. Arbitrary changes of task or interface are outside the evaluated scope.

  • Budgets are matched by rounds and candidates

    Equal optimization rounds do not establish equal token use or wall-clock cost. SIM is read in full, so scaling the memory has a cost.

  • Continuation can lose

    One coding pair favors no SIM. In the Gemini generation chain, CRSI also finishes 0.8 points below fresh RSI on Gemini 4 Argon (Table 1).

  • Preprint, not peer-reviewed

    The results are reported by the authors. Independent source-run repetition is concentrated in the dedicated ten-pair coding study.

06 / Read & discuss

Follow the work.

The paper contains the full method, settings, case studies and result tables. This page is a companion to the November 13, 2026 preprint.

01 / THE PAPER

The full study

Self-Improvement Memory Enables Continual Recursive Self-Improvement

arXiv ↗   PDF ↗
02 / THE CONVERSATION

Questions & perspectives

Discuss how optimization experience can carry forward when agent systems move to new models or benchmarks.

Hugging Face Papers ↗
CITE THIS WORK
@misc{yu2026selfimprovementmemory,
  title={Self-Improvement Memory Enables Continual Recursive Self-Improvement},
  author={Xin Yu and Fei Xia and Jiaqi Pan and Youbao Tang and Chengzhi Liu
          and Adams Yu and Sicong Jiang and Yanfei Chen and Lingzhou Xue and Ting Yu},
  year={2026},
  eprint={2611.18856},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2611.18856}
}

No public implementation is linked here. The interactive record is illustrative; charts reproduce reported paper values. Download the visual overview ↗