memory-centric-agentic-research

# Memory-Centric Agentic Research ## Overview Automated scientific research requires persistent memory across projects, structured lifecycle execution, and self-improvement from feedback. AutoSci (arXiv:2605.31468) provides a unified four-module architecture for the full research lifecycle: literature understanding → hypothesis formulation → experimentation → manuscript writing → rebuttal. ## Core Modules ### SciMem: Schema-Governed Research Memory Two-tier memory separation: - **Long-Term Knowledge Memory (LTKM)**: Reusable scientific knowledge that persists across all projects - **Active Research Memory (ARM)**: Project-level artifacts (ideas, experiments, manuscripts, reviews) **Key principle**: Schema governance ensures structured, queryable memory rather than unstructured text dumps. Memory entries follow consistent schemas for retrieval and reasoning. ### SciFlow: Five-Stage Lifecycle Execution Controlled harness manages: 1. **Literature Understanding** — parse and synthesize related work 2. **Hypothesis Formulation** — generate testable hypotheses from literature gaps 3. **Experimentation** — design and execute experiments 4. **Manuscript Writing** — produce publication-quality papers 5. **Rebuttal** — respond to reviewer feedback Each stage uses state management, context preservation, verification, and feedback orchestration. ### SciDAG: DAG-Shaped Multi-Agent Operators For complex skills requiring coordination: - Represent difficult operations as DAGs of sub-tasks - Reuse stage-specific templates across research projects - Enable parallel execution where dependencies allow ### SciEvolve: Self-Evolving System Converts feedback signals into versioned updates: - **User feedback** → update SciMem organization - **Experiment results** → refine SciFlow skills - **Review responses** → improve SciDAG templates - **External environment** → adapt all components ## Methodology ### Designing a Persistent Research Agent 1. **Define memory schema** — What knowledge persists vs. what is project-scoped? 2. **Implement lifecycle harness** — State machine controlling research stages with verification checkpoints 3. **Build multi-agent DAGs** — Decompose complex operations (e.g., "run experiment") into coordinated sub-tasks 4. **Add evolution loop** — All feedback must produce versioned updates, not just logging ### Key Design Decisions | Decision | AutoSci Approach | Why | |----------|-----------------|-----| | Memory organization | Two-tier (LTKM + ARM) | Cross-project knowledge vs. project artifacts have different lifecycles | | Search structure | Explicit parent pointers (tree) | Implicit trace history loses structural information during backtracking | | Skill evolution | Versioned updates from feedback | Ensures traceable improvement, prevents regression | ### Integration with Existing Agent Systems - **Compatible with**: Any LLM-based agent framework with tool-use capability - **Requires**: Persistent storage (database), state machine controller, feedback collection mechanism - **Extends**: Single-turn agents → persistent multi-project research environments ## Pitfalls - **Memory bloat**: Without schema governance, research memory becomes unstructured and useless. Enforce strict schemas for all memory entries. - **Stage coupling**: Research stages must be independently verifiable. If one stage fails, later stages should not silently produce incorrect results. - **Feedback overload**: Not all feedback signals are equally useful. Prioritize experiment results and reviewer feedback over environmental noise. - **Template rigidity**: SciDAG templates should be parameterized, not hardcoded. Overly rigid templates prevent adaptation to novel research scenarios. ## Related Skills - `agent-coordinator` — Agent orchestration patterns - `autopoiesis-self-evolving-systems` — Self-evolving system paradigms - `coral-open-ended-discovery` — Open-ended discovery workflows