AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle

Paper metadata

| Attribute | Content | |---|---| | Title | AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle | | Authors | Weitong Qian, Beicheng Xu, Zhongao Xie, Bowen Fan, Guozheng Tang, Jiale Chen, Xinzhe Wu, Mingtian Yang, Chenyang Di, Jiajun Li, Lingching Tung, Peichao Lai, Yifei Xia, Ziyi Guo, Yanwei Xu, Yanzhao Qin, Shaoduo Gan, Xupeng Miao, Bin Cui | | Institution | Peking University | | arXiv ID | 2605.31468 | | Date | 2026-05-29 | | Category | cs.AI | | Core claim | Automated research should not be a patchwork of scattered tools, but a unified system with structured persistent memory, full lifecycle execution, and system-wide self-evolution |

1. A PhD student's five years, compressed into a system

Consider a new PhD student. Year one: reading literature, drawing concept maps. Year two: generating ideas, running experiments, failing, retrying. Year three: obtaining credible results and writing a paper. Year four: responding to reviews and revising. Year five: publishing and starting the next project with accumulated knowledge.

What accumulates is not a stack of papers but structured knowledge — typed entities (papers, concepts, methods, researchers) and explicit relationships stored in long-term memory, not scattered PDF annotations.

AutoSci's ambition is to compress these five years into an executable system. Its answer consists of four modules: SciMem (memory), SciFlow (workflow), SciDAG (multi-agent augmentation), and SciEvolve (evolution). These form a closed loop: memory provides substrate for the workflow, the workflow produces artifacts written back to memory, multi-agent augmentation handles complex stages, and evolution converts feedback into system updates.

The real gap in prior work was integration: existing systems either covered part of the lifecycle, had full loops but memory reset per project, or accumulated experience without modifying themselves.

2. SciMem: research memory as a typed knowledge graph

SciMem divides memory into two regions by lifecycle and reuse scope.

2.1 Long-Term Knowledge Memory

Stores reusable scientific knowledge as six typed entity types:

• Topic: domain scope and key observations
• Paper: structured reading notes
• Foundation: stable background knowledge
• Concept: reusable scientific ideas or terminology
• Method: reusable technical methods with implementation details
• People: researcher profiles and fields

Relations are mechanically checkable explicit links (e.g., Paper–[introduces/critiques]→Concept, Paper–[applies/extends]→Method). The memory is semantically addressable and incrementally extendable.

In the GPU kernel optimization case, the Topic entity "LLM-Based Kernel Generation" includes a timeline, milestone works, a SOTA tracker (e.g., frontier reasoning models below 20% correctness on KernelBench; kernelfoundry at 97% correctness with 2.32x speedup; AscendOptimizer at 1.21x geomean speedup on 101 AscendC operators), and open problems (72% fusion-task failure rate across methods; 46.6% of correct kernels slower than baseline; cross-hardware speedup variance up to 21.4x).

2.2 Active Research Memory

Tracks project-level state with lifecycle statuses for Idea, Experiment, Manuscript, and Review. Terminal artifacts flow back into long-term memory as a consolidation bridge.

2.3 Memory growth and Trust Guard

Memory expands via long-horizon aggregation, cross-region flow, and cross-cycle accumulation. All writes pass Trust Guard: deterministic linting (schema fields, lifecycle states, link types, bidirectional links) plus independent reviewer-agent content checks (evidence support and consistency). Writes are marked pass, warn, or block; blocked artifacts are quarantined. This guards against memory pollution producing cascading false conclusions.

3. SciFlow: five stages and an execution harness

The lifecycle has five stages — Literature, Ideation, Experiment, Writing, Rebuttal — each implemented as harness-based skill contracts with explicit memory reads/writes. The harness provides five guarantees:

• State: persistent storage of stage outputs, lifecycle states, links, and pipeline progress, enabling resumption from any stage
• Context: each skill receives a custom SciMem view rather than the full memory graph
• Verification: Trust Guard checks memory writes and high-risk handoffs
• Feedback: failures and critiques trigger /refine or self-evolution
• Orchestration: /research loop with non-blocking execution and monitoring for long-running experiments

SciFlow implements 30+ structured research skills with defined inputs, required context, checkpoints, output formats, and handoff rules.

4. SciDAG: DAG-shaped multi-agent augmentation

For stages needing broader search, debate, or refinement, SciDAG invokes operators (generate, variation, debate, refine, review — 9 reusable operators) as nodes in a directed acyclic graph, with conditional edges routed by quality, cost, and convergence signals. Common graphs are stored as stage-aware templates with execution traces written back for reuse.

The debate operator pits opposing sub-agents against each other along defined dimensions (novelty, feasibility, impact, relation to prior work), producing comparative analyses rather than winner-take-all outcomes. The refine operator diagnoses risk sources before targeted revision, drawing on stored failure experience.

5. SciEvolve: system-wide self-evolution

SciEvolve collects signals from user, task, and open environments, then applies three evolution paths:

• /dream — SciMem evolution: periodic review of traces and memory neighborhoods; demotes/archives stale entries, compresses redundancy, merges entities, proposes new cross-entity links
• /forge — SciFlow evolution: treats skills as versioned research protocols; analyzes repeated failures and user corrections, then proposes auditable patches (e.g., strengthening claim-evidence checks in writing skills)
• /morph — SciDAG evolution: uses traces to revise operator prompts/roles/tools, prune weak branches, add verification nodes, or specialize templates

The key insight: true self-improvement is not accumulating reusable text experience, but converting repeated feedback into controlled updates of the system's own structure.

6. Case studies

Case 1: GPU kernel optimization (6.3/10)

Starting from seed papers, AutoSci built long-term memory, generated five candidate directions, killed duplicates via /novelty, screened under a 4×A40 budget via /exp-pilot-run, and selected a profiler-guided Claude Code agent approach in a TritonBench workspace. The resulting paper-level artifacts scored 6.3/10 on automated ICLR-style review.

Case 2: Biomedical drug discovery (5.8/10) — valorizing negative results

For structure-aware post-translational modification (PTM) modeling, five candidates were proposed; novelty checks removed two (saturated subspace; non-fine-tunable AF3 weights), and scoring prioritization selected candidate E (a PTM-aware degrader target nomination) for execution on an RTX 4060.

The core premise was falsified: across 15 POIs / 189 interface residues, phosphorylation (14.5%), alanine scanning (15.9%), and Kme3 (15.7%) barely exceeded the 13.4% random level (p>0.3, effect-size CIs including 0, 0/69 surviving BH-FDR/Bonferroni correction). Dose-response controls confirmed the operator was not inert — the bottleneck was the scorer's dynamic range.

AutoSci converted the negative bound into a preregistered benchmark — a threshold any future PTM-sensitive scorer must clear — and combined it with a deferred idea to generate a follow-up plan: a PTM-sensitive ternary scorer. Automated ICLR review score: 5.8/10.

7. Evaluation and limitations

The scores measure the reviewability of auto-generated artifacts, not superiority over top human papers. More significant is the negative-result valorization mechanism: instead of forgotten footnotes, negative findings become explicit constraints in long-term memory, countering publication bias and saving future researchers from dead ends.

Acknowledged limitations:

• Compute cost: each full cycle consumes heavy token and compute budgets
• Domain dependence: experiments are mostly computational; wet-lab agents remain future work
• Quality ceiling: 5–6 review scores indicate "acceptable research," not breakthroughs
• Memory scale: retrieval efficiency and consistency at very large graph sizes are unverified
• Human-machine interface: command-driven workflows (/research, /ideate, /forge) pose a barrier for non-technical researchers
• Ethics and accountability: responsibility for errors in published AI-generated work remains a governance question, though harness audit logs provide traceability

8. Positioning vs. related work

| System | Full lifecycle | Persistent memory | System-wide evolution | |---|---|---|---| | EvoScientist | ✓ | ○ | — | | DeepScientist | ✓ | — | — | | ARIS | ✓ | ○ | — | | NORA | ✓ | ○ | — | | Deep Researcher Agent | ✓ | ○ | — | | AutoSci | ✓ | ✓ | ✓ |

(✓ = fully supported, ○ = partial or project-local, — = not a primary focus)

AutoSci uniquely combines persistent memory with whole-system evolution: memory serves not just as an experience repository but as the substrate for self-update.

9. Conclusion

AutoSci proposes not a smarter research assistant but a persistent research environment that executes projects, remembers experience, and evolves across projects. The bottleneck of automated science is not the quality of individual skills but how skills hand off, how memory accumulates across projects, and how systems learn from failure. The scores of 6.3 and 5.8 counsel against over-optimism — AutoSci cannot yet replace human researchers — but it redefines the baseline question: not "can AI do one step," but "can AI complete the full journey and keep improving." The answer is moving from "no" toward "just beginning."

References

1. Qian, W., Xu, B., Xie, Z., et al. (2026). *AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle*. arXiv:2605.31468 [cs.AI]. 2. Lyu, S., et al. (2026). *EvoScientist: Evolving Scientific Research via Large Language Models*. arXiv preprint. 3. Weng, Y., et al. (2025). *DeepScientist: Advancing Scientific Discovery through LLM Agents*. arXiv preprint. 4. Yang, S., et al. (2026). *ARIS: Autonomous Research and Intelligence System*. arXiv preprint. 5. Zhou, J., et al. (2026). *NORA: Neural Orchestrated Research Assistant*. arXiv preprint.