Source-linked AI summary
The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
Yi Duan, Ying Liu, Zirui Tang, Haodong Chen, Jun Zhou, Yumou Liu, Bangrui Xu, Yukai Wu, Sidi Chen, Yuhan Zhou, Haoyu Wang, Xiaoyou Yu, Shaokun Han, Xuzhou Zhu, Le Zhou, Bolin Lu, Wei Zhou, Jiachen Liu, Nuozhou Fang, Jiaxin Tian, Ruoyu Chen, Yuxuan Li, Kai Zuo, Kaiyan Zhang, Jiantao Qiu, Conghui He, Guoliang Li, Bowen Zhou, Zhiyuan Liu, Zhoufutu Wen, Jihua Kang, Xuanhe Zhou, Fan Wu
TL;DR
The paper addresses the bottleneck of scaling AI improvement while existing systems struggle to retain experience and autonomously coordinate improvement. It proposes a five-stage autonomy-centered RSI roadmap, connects it to applications and industrial systems, and identifies evidence-backed progress alongside unresolved requirements for genuine RSI.
Problem
Scaling AI improvement is bottlenecked because developers must still choose improvements, build resources, and verify changes, while current systems often discard experience after tasks end.
Method
The paper develops an autonomy-centered RSI roadmap spanning improvement execution, strategy selection, experience acquisition, environmental adaptation, and recursive meta-improvement, grounded in application scenarios and industrial practices.
Results
The paper shows uneven capability progress and reports industrial RSI systems that automate iterative improvement, including ForgeTrain surpassing Megatron-LM v0.15 within 1.5–2.5 days.
Takeaways & Limitations
Progress toward genuine RSI depends on persistent learning, reliable feedback and attribution, human-governed validation, and adaptation to domain-specific deployment constraints.
Takeaways & Limitations
Automated pipelines do not by themselves establish experience autonomy without closed dependence among learner state, experience acquisition, persistent learning, and later acquisition decisions.
Abstract
from arXiv · showhide
Recursive self-improvement (RSI) enables AI systems to turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement. We first use the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, then introduce the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. Next we examine RSI across scenarios (e.g., scientific discovery, embodied intelligence, software engineering), highlighting their distinct requirements and development speeds. Drawing on diverse industry practices and preliminary empirical evidence, we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.
1 Introduction
The paper frames recursive self-improvement as a response to costly, externally coordinated development, defining it as a closed loop that improves both AI capabilities and the improvement process. It organizes RSI by transferred responsibility, distinguishes genuine recursion from performance gains alone, and illustrates the framework through training and software-engineering systems.
- From Development Burden to RSI: RSI is an autonomous closed loop in which AI identifies limitations, develops and validates improvements, and uses resulting capabilities to improve future improvement.The paper presents RSI as a response to development burdens created by externally coordinated interventions.
- From Development Burden to RSI: RSI spans autonomy, efficiency, and innovation, extending beyond task performance to improve how later improvements are discovered and implemented.The framework emphasizes persistent changes that alter subsequent improvement rounds rather than one-off optimization.
- Illustrative RSI Systems: A-Evolve-Training retains post-training outcomes in a research policy and discovery log that guides later recipe choices, while Ouroboros updates tools, context, prompts, and implementation after reviewed deployment evidence.Both cases modify the process used for later improvement, not only the immediate model or software product.
- Challenges for RSI: Safe inheritance, autonomy attribution, and reliable verification remain central challenges because persistent updates can degrade capabilities, leave selection rules external, or exploit evaluators.The paper highlights transfer tests, rollback, protected evaluation, independent anchors, and matched budgets as safeguards.
- An Autonomy-Centered Framework: The roadmap transfers responsibility across execution, strategy selection, experience acquisition, environmental adaptation, and recursive inheritance.These levels track which improvement decisions AI assumes and which remain externally controlled.
- Evidence and Challenges: RSI evidence must separate structural recursion from effective recursion because higher task performance alone does not show that an improvement mechanism was revised, retained, and reused.The paper also requires distinguishing AI-controlled decisions from fixed search procedures and human acceptance criteria.
2 Background and Preliminaries
This section combines HCI-based evidence of uneven capability progress with an autonomy-centered definition of recursive self-improvement. It frames RSI as a persistent improvement loop that generates, validates, retains, and inherits changes, while drawing on industrial practice to analyze contemporary systems.
- 2.1 Uneven Capability Progress in Modern Foundation Models: The HCI normalizes benchmark-family scores from entry-year frontiers to perfect-score endpoints and aggregates eligible benchmark families using coverage-sensitive weighting.The square-root weight increases the contribution of better-covered families without allowing the largest table to dominate the domain value.
- 2.1 Uneven Capability Progress in Modern Foundation Models: By 2026, capability levels and trajectories diverge sharply across domains, with advanced mathematics and graduate-level science far ahead of interactive capabilities such as tool agents.Advanced mathematics reaches 86.4 and graduate-level science 85.8, while software engineering reaches 52.6 and tool agents 39.9.
- 2.1 Uneven Capability Progress in Modern Foundation Models: The illustrative post-2026 RSI extension assigns larger gains to domains with more unclosed headroom, moving software engineering from 52.6 to 89.6 and tool agents from 39.9 to 86.8.The extension uses Rd−Td,2026 = 0.78(100−Td,2026); cybersecurity changes from 91.9 to 98.2 under the same illustration.
- 2.2.1 Anatomy of an Improvement Loop: An improvement loop repeatedly uses experience to propose and evaluate a modification, retains accepted changes, and starts the next round from the updated state.The analysis asks what generates experience, what changes, what persists, who controls updates, and whether one round shapes later rounds.
- 2.2.2 Definition of Recursive Self-Improvement: RSI is defined as an autonomous capability that converts experience and feedback into persistent changes that can alter later improvement mechanisms.The improved system re-enters subsequent interaction and improvement rounds with a changed capability state.
- 2.3 Scope of Evidence: Industrial practice is included because technical reports, repositories, blogs, model releases, and deployed infrastructures often expose operational details of improvement loops before academic publication.These sources reveal retained artifacts, evaluator and tool interactions, and whether improvement mechanisms continue beyond a single experiment.
3 RSI Across Autonomy Levels
The paper organizes RSI as a hierarchy of increasing autonomy, from in-task refinement without persistent system change to systems that select improvement strategies under externally fixed objectives and evaluation. Across levels, it tracks where the loop closes, what persists, and which decisions remain human-controlled.
- The hierarchy progresses from in-session refinement toward recursive improvement, where mechanisms that produce future improvements also become subject to improvement.The framework classifies existing work by how much responsibility AI assumes for its own improvement.
- B0: In-Task AI Improvement: B0 is a non-RSI reference level: systems improve current outputs, but retain no resulting change as persistent state for future tasks.Examples include iterative revision, reflection, or search confined to the current session.
- B0: In-Task AI Improvement: B0 systems cannot accumulate experience across tasks, so corrections remain temporary and errors may recur despite additional iterations.Intermediate results, critiques, and corrections do not become persistent parameters, reusable memory, or operating rules.
- L1: Improvement Execution Autonomy: At L1, AI executes human-defined improvement procedures and validation rules, while resulting artifacts persist and enter later workflows.L1 evaluation therefore concerns execution reliability and the safety of retaining changes for subsequent development.
- L2: Improvement Strategy Autonomy: At L2, AI diagnoses weaknesses and chooses improvement interventions, but humans still specify objectives, task boundaries, and acceptance criteria.This shifts the bottleneck from executing known improvements to searching over possible improvements without granting authority to redefine success.
- L2: Improvement Strategy Autonomy: Across autonomous search, repeated evaluator access can produce benchmark overfitting, shortcut discovery, or other apparent gains unrelated to genuine capability improvement.The paper identifies evaluator blind spots, search-compute confounding, and test-label extraction as concrete hazards.
3.4 L3: Autonomy over Future Learning Experience
L3 grants the system autonomy over what learning experience to acquire next, using evolving learner evidence to steer selection and feeding persistent learning back into future acquisition. The section distinguishes this closed loop from merely automating data operations and surveys adaptive task generation, self-play, and autonomous practice.
- L3: Autonomy over Future Learning Experience: The characteristic L3 loop observes learner state, chooses and acquires experience, updates persistent state, reshapes future experience, and repeats.
- From Industrial Automation to Experience Autonomy: Automated ingestion, filtering, labeling, generation, and orchestration do not establish experience autonomy without closed dependence between learner feedback, acquisition, persistent learning, and later acquisition.Human-designed selection rules can still support L3 when they respond to changing learner feedback; one-time filtering or retention is insufficient.
- Autonomous Practice through Environment Interaction: Autonomous practice selects goals from current skills and environmental observations, consolidating successful interaction experience into reusable skills.Figure 5 presents adaptive task generation and self-play alongside environment-based practice as complementary realizations of learner-conditioned acquisition.
- Adaptive Task Generation and Self-Play: Adaptive task generation targets the learner’s changing competence frontier, while self-play and challenger–solver systems use learner-dependent signals to evolve future tasks.Consistency-based difficulty estimates can make acquisition responsive when correctness signals are unavailable, but they do not independently establish correctness or future learning gain.
- Evaluating L3 Experience Autonomy: SIMA 2’s fixed-task self-generated trajectories show improvement but do not alone establish task-selection autonomy, unlike ASKA’s practice allocation based on weaker skills.
- L3: Autonomy over Future Learning Experience: L3 uses learner capabilities, failures, or learning history to select, generate, or seek subsequent experience, while persistent learning reshapes later acquisition decisions.Human-designed objectives, evaluators, and learning rules may remain in place; the defining feature is the feedback loop between learner state and future experience.
3.5 L4: Autonomy in Deployment and Environmental Adaptation
L4 shifts autonomy from choosing learning experience to deciding which consequences of deployment experience persist in the operating agent. The section organizes adaptation around trajectory distillation, agent revision, selective retention, evaluation, and the risks of unreliable long-term reuse.
- L4: Autonomy in Deployment and Environmental Adaptation: The L4 loop observes deployment interaction, proposes persistent adaptation, revises agent components, validates and retains updates, reuses them, and repeats with new feedback.
- Trajectory Distillation: Trajectory distillation makes interaction histories reusable by converting them into compact textual memories, structured records, procedural skill libraries, or executable artifacts.These artifact types differ in how experience is represented and reused across later tasks.
- Trajectory Distillation: Executable artifacts combine textual plans and environment knowledge with callable tools, while evaluation protocols measure long-horizon reuse, within-run behavior, or execution and construction costs.
- Iterative Revision of the Agent System: Useful persistent updates and the ability to exploit them at solve time are distinct, because failures arise when relevant artifacts are not activated or not faithfully followed.The reviewed analysis finds updating ability largely independent of base-model capability, while benefit from updated harnesses varies non-monotonically.
- Selective Retention and Deployment: Unbounded skill accumulation can degrade retrieval quality and stall progress, motivating lifecycle management based on evidence, retirement, and maintenance of retained libraries.
- Limitations and Risks: Long-term deployment adaptation remains uncertain because evaluations are usually bounded, feedback can be noisy or endogenous, and retained artifacts may fail to influence execution.
- L4: Autonomy in Deployment and Environmental Adaptation: L4 converts deployment interaction histories into retained changes to memory, skills, harnesses, code, or model parameters that affect later tasks and feedback.Humans retain objectives, protected acceptance criteria, access boundaries, and final authority over consequential releases.
3.6 L5: From Environmental Adaptation to Meta-Improvement
L5 makes the mechanism responsible for future improvement an inheritable target, allowing revised improvers, evaluators, or research policies to govern later rounds while humans retain external objectives and controls. Evidence supports bounded meta-improvement, but reliable accumulation across generations remains unestablished.
- L5 begins when AI persistently revises a mechanism governing future improvement and reuses it to generate, evaluate, or select later successors.The inherited state can include code, prompts, evaluators, research policies, and supporting evidence, while human designers retain mission, protected evaluation, and resource authority.
- Structural L5 shows persistent control of a later improvement round, whereas effective L5 additionally requires better successors under comparable budgets and independent assessment.Self-modifying task code alone is insufficient when the later-revision process remains unchanged.
- 3.6.1 Improving the Search Procedure: The search procedure can itself be improved: STOP selected a fourth-generation improver that outperformed the seed on all five transfer tasks excluded from self-improvement.The task distribution, utility, base model, and resource limits remained externally specified; weaker-model runs regressed on average, and some programs exploited evaluation weaknesses.
- 3.6.2 Improving Successor Evaluation: RQGM co-evolves successor evaluators while freezing each evaluator within an epoch and comparing challengers against an independent ground-truth anchor.On held-out Polyglot coding tasks, the preprint reports a 71.7% pass rate versus 69.9% for HGM-H, with lower search-token use.
- 3.6.3 Revising Research Goals and Policies: A-Evolve-Training revises a persistent research policy and discovery log across four post-training rounds, directing later experiments when development scores diverge from external gains.The study reports a final leaderboard score of 0.86 versus 0.87 for the top human submission, while the objective and worker substrate remain human-defined and fixed.
- 3.6.5 Evaluating L5 Recursive Improvement: Current evidence supports bounded meta-improvement, but statistically reliable accumulation across generations under comparable resources remains open.Safe inheritance requires transfer tests, version histories, rollback, protected evaluation, and matched computational budgets to distinguish useful updates from regression or evaluator exploitation.
4 RSI Across Applications
RSI applications differ in the components they improve, the feedback needed to validate updates, and how far current systems progress toward persistent, system-level evolution. Across science, embodied intelligence, software engineering, and healthcare, current work reaches persistent updates and selective autonomy, but remains partial and domain-constrained.
- AI for science: Scientific RSI must improve hypothesis generation, experiment execution, and evidence interpretation, not merely produce better individual discoveries.Open-ended search, sparse and delayed feedback, and uncertainty make evidence acquisition and epistemic control central requirements.
- AI for science: Current AI4Sci systems span widespread output refinement, demonstrated persistent updates, and a frontier of autonomous updates to weights, tools, skills, workflows, and scaffolds.A few systems also show aspects of adaptive exploration and cross-task experience reuse.
- Embodied intelligence: Embodied RSI should improve both task performance and capability acquisition by coordinating environment generation, harnesses, policies, world models, evaluators, and inheritance.Verified improvements can be inherited by the policy, world model, or harness so interaction cycles improve future learning as well as current behavior.
- Software engineering: Software-engineering RSI forms a loop from development challenge through agent modification, repository-grounded trial, regression-aware selection, and versioned inheritance.Full L5 improvement of the software-development improver has not yet been demonstrated, while environment adaptation remains largely absent.
- Healthcare: Healthcare RSI requires persistent task improvement while preserving clinical validity and safety, with higher levels constrained by longitudinal attribution, safe feedback acquisition, and selective transfer.Every persistent update should be clinically validated, monitored, traceable to supporting evidence, and reversible when assumptions or safety guarantees fail.
5 Industry Landscape and Preliminary Practices
Industry practices instantiate RSI through environment–data–model co-evolution, automated evaluation, versioned quality-system updates, autonomous engineering, and executable research artifacts. Preliminary results show measurable gains, while the systems primarily demonstrate components or bounded forms of recursive improvement rather than complete RSI.
- Theseus: Theseus co-evolves environments, data, and models by refining environments, generating targeted tasks, training task models, and iterating toward further tasks.The loop converts refinement experience into environment-refinement tasks and uses reconstructed environments to expose genuine task-solving difficulties.
- Theseus: Clean workspaces raised pass rates by 21.7–51.6 percentage points across eight model–harness configurations on 30 workspace tasks.The pilot used 1,280 rubrics and compared clean with noise-laden workspace conditions.
- Humanlaya: Humanlaya promotes reviewed, versioned quality-system modifications using delivery evidence and held-out task packages, rather than treating current-artifact improvement alone as persistent RSI.After four updates, key-defect packages fell from 9.0% to 3.7% and average human handling time from 48 to 27 minutes.
- ModelBest and Hyra: Forge Engineering and Hyra extend autonomous improvement to implementations and evaluators, with reported gains across training systems, kernels, and AI-for-AI and AI-for-Science tasks.ForgeTrain reportedly surpassed Megatron-LM v0.15 within 1.5–2.5 days, while Hyra reports validation BPB of 0.9015 versus 0.9109 on nanochat AutoResearch.
- Agent-Native Research Lab: Agent-Native Research Lab proposes executable research artifacts that preserve logic, code, exploration branches, and raw evidence so successors can inherit verifiable research experience.Its ARA evaluation reports 93.7% question-answering accuracy over prior work using the artifact.
6 Challenges and Future Directions
The central challenge is proving that inherited changes improve the capacity to improve across later rounds, rather than merely producing retained artifacts or one-off gains. Future RSI systems therefore need attributable diagnosis, reliable learning signals, validated reuse, domain-specific governance, trustworthy evolving evaluators, and resource-aware long-horizon evaluation.
- Diagnosis and coordinated improvement: Cross-component diagnosis is difficult because failures may originate in data, context, tools, models, evaluators, or their interactions.Future systems should test intervention hypotheses with controlled comparisons, targeted ablations, and dependency records before coordinating changes across components.
- Experience acquisition: Learner-conditioned experience must preserve validity, difficulty, diversity, and coverage while estimating whether experience produces durable learning value.Correctness or estimated difficulty alone does not establish that an experience improves later learning.
- Persistent-state management: Persistent artifacts require evidence, applicability conditions, dependencies, activation, faithful use, and downstream benefit; retention alone does not establish useful reuse.Validation must continue as task distributions, executors, and surrounding systems change.
- Governed adaptation: Domain-specific feedback constrains autonomous learning and transfer because scientific, embodied, and healthcare outcomes differ in observability, cost, uncertainty, delay, and population dependence.Healthcare updates may not transfer across populations or institutions, and deployment requires governed, traceable, reversible adaptation.
- Trustworthy evolution: Evolving improvers and evaluators create a coupled validation problem because stronger solvers can expose evaluator weaknesses while evaluator changes can make scores incomparable or exploitable.Suggested safeguards include frozen evaluators within epochs, independent anchors, and revisiting scores affected by replacement.
- Evaluation and resources: Long-horizon evaluation must compare original and revised mechanisms under matched budgets and report trajectories including rejected updates, regressions, recovery, retained capabilities, and resource costs.Structural inheritance is insufficient; effectiveness requires better subsequent improvements.
- Core evaluation challenge: Genuine RSI requires repeated, attributable, and transferable improvements in the capacity to improve under explicit resource and authority constraints.Evidence must show that inherited changes help later rounds acquire experience, discover interventions, or validate successors more effectively.
7 Conclusion
The report presents RSI as an autonomy-centered roadmap spanning improvement execution, strategy selection, experience acquisition, environmental adaptation, and recursive meta-improvement. It connects these stages to domain-specific evidence and argues that long-horizon evaluation is needed to establish transferable gains under resource constraints and human oversight.
- Conclusion: The roadmap spans five RSI stages: improvement execution, strategy selection, experience acquisition, environmental adaptation, and recursive meta-improvement.It uses the improvement loop to clarify what systems change, what successors inherit, and which decisions remain human-controlled.
- Conclusion: Progress toward RSI depends on domain-specific feedback availability, verification costs, deployment constraints, industry practices, and preliminary empirical evidence.The report emphasizes that long-horizon evaluation must test transferable gains under explicit resource constraints and human oversight.
Appendix A: RSI Landscape
The RSI landscape organizes surveyed systems by autonomy level and by the components or targets they improve. The taxonomy spans memory, workflows, tools, models, training, evaluation, environments, artifacts, and full-system improvement loops.
- Figure 16 summarizes surveyed papers by autonomy level, primary improvement target, and sub-target across 491 papers.Autonomy percentages are based on paper counts, while papers with multiple targets contribute fractional weights.
- The taxonomy covers memory, workflow and control, tools, models, training, evaluation and feedback, data and environments, outputs, and full-system co-evolution.Table 11 maps each sub-target to the concrete system component subject to modification.
- Improvement targets include persistent memories, execution graphs and planning, routing and scheduling, tool interfaces and implementations, skills, model parameters, architectures, and inference policies.The framework distinguishes surrounding harness components from the backbone model and its runtime configuration.
- Training and feedback targets include objectives, optimizers, schedules, curricula, search procedures, evaluators, reward functions, verifiers, and credit assignment.These components generate, assess, select, and convert outcomes into signals for candidate improvements.
- The taxonomy also includes task and curriculum generation, simulators and world models, produced programs and scientific artifacts, and mechanisms for generating, evaluating, selecting, and applying changes across cycles.Full-system categories include multi-target harnesses, model–harness coevolution, and improvement-loop or meta-RSI mechanisms.
Appendix B: Industry Landscape
The industry landscape is organized by company archetype and improvement target.
- Table 12 organizes surveyed industrial systems by company archetype and improvement target.
Company Archetypes in the Emerging RSI / Self-Improvement Landscape
The emerging industry landscape groups representative products and research systems by company archetype, while mapping their self-improvement targets and RSI relationships. These systems span agent training, reasoning, computer use, persistent agents, model customization, and verification.
- The landscape contains 72 distinct companies or teams in a public-source snapshot from September 2026.Each row represents one product or representative work.
- The organization principle is company archetype rather than geography, and the RSI tag maps systems to the B0–L5 scheme without claiming that companies use that label.
- Other systems improve agent harnesses, code generation and verification, research pipelines, training workflows, and self-editing or self-evolving agents.Examples include critic–revise loops, verifier-based systems, research automation, and self-editing agents.
- Systems cover agent training, synthetic environments, reasoning-policy bootstrapping, computer-use agents, persistent memory and skills, orchestration, and model customization.The listed AI-controlled components include policies, weights, environments, memories, skills, and tools.
B. RSI-native / AI4AI companies
RSI-native and AI4AI companies center their work on self-improving research, agents, code, models, and infrastructure. Their systems range from automated research workflows and shared code evolution to candidate L5 meta-improvement visions.
- The RSI-native category treats self-improvement, AI-for-AI, self-evolution, or recursive improvement as central to the company or research thesis.
- Other efforts target AI systems and chips, model or harness components, training recipes, datasets, code, and research policies through search, evaluation, and iteration.
- The landscape includes candidate L5 visions involving improver rewrites, inheritance, self-modifying research pipelines, and meta-improvement of research agents.
- Several systems evolve agents through experience, memory, environmental tasks, feedback, refactoring, and skill reuse.
- Representative systems automate research through hypothesis generation, experiment composition, code execution, scoring, and iterative refinement.Examples span AI research automation, autoresearch workflows, and self-evolving research systems.
C. Autonomous R&D & scientific-discovery companies
Autonomous R&D systems automate parts of scientific discovery, from hypotheses and experiments to analysis, formal proof, and research reporting. The listed systems span narrow task automation and broader research workflows, mostly at L1–L2, with some L2–L3 candidates.
- Formal-mathematics systems focus on conjecture, proof search, verifier traces, and formal proof validation, while research agents extend workflows toward experiments, coding, and reports.
- Scientific-discovery platforms automate hypothesis generation, experiments, data collection, and analysis across laboratory and computational workflows.
- Research-model systems automate training-data design, code or configuration changes, GPU runs, benchmarking, ablations, and iterative model improvement.
D. Agent optimization, evaluation & learning infrastructure
Agent-optimization infrastructure targets the feedback, evaluation, memory, data, and reinforcement-learning loops required for persistent system improvement. The ecosystem includes tools for skill rewriting, evaluator updates, simulation, production-feedback mining, and training-environment construction.
- The infrastructure layer supplies feedback, skills, reinforcement learning, evaluation, simulators, training data, and production-agent improvement capabilities.
- Optimization systems mine failures or feedback to update skills, prompts, evaluators, memories, or agent behavior through repeated measurement and revision.
- Evaluation and learning infrastructure generates scenarios, grades behavior, ingests traces, constructs datasets, and supports distributed reinforcement-learning environments.
E. Embodied, world-model & continual-adaptation companies
Embodied and continual-adaptation systems connect model improvement to world models, robotics, deployment, and test-time learning. Their improvement targets include embodied policies, tactile models, action data, environment modeling, and adaptation harnesses.
- Embodied platforms combine VLA models, data, training, evaluation, and real-robot deployment for robotics research and adaptation.
- Tactile and world-model systems target multimodal sensing, prediction, manipulation policies, environment modeling, and skill acquisition.
- Continual and test-time learning systems use mechanisms such as LoRA, skill vectors, fast weights, and adaptation harnesses to support context adaptation and skill transfer.
F. Persistent-memory & personal-AI companies
Persistent-memory and personal-AI systems treat memory, shared experience, reusable skills, and identity as substrates for adaptation. The catalog includes knowledge modules, long-term agent memory, evolving local agents, and personalization systems, with some entries marked as requiring stronger verification.
- Cross-task memory, identity, and reusable skills provide the main substrate for persistent adaptation in this system category.
- Agent-memory systems store, retrieve, and adapt knowledge or skills through persistent memory modules, skill forges, and long-term memory operating systems.
- Personal-agent systems combine workspace persistence, memory, experience libraries, identity models, and personalization for ongoing user adaptation.
- The catalog distinguishes task-local autonomy and RSI-adjacent or candidate capabilities, while some company identities or technical boundaries require stronger verification.