File System-First Agentic Research Framework: Enabling AI Agents to Have Persistent Research Memory

This introduces the Agentic Research Harness framework, whose core is to persist research states to the file system instead of relying on conversation context, solving issues like context loss and difficulty in source tracing in AI research. The framework emphasizes "one folder per problem, goals first, separation of sources and synthesis, and binding claims to evidence", ensuring the persistence, auditability, and collaboration of research through structured file storage.

With the development of large language models, AI-assisted research has become popular, but traditional workflows use chat conversations as the workspace, leading to the loss of context, discoveries, and conclusions after conversations end. The Agentic Research Harness framework proposes a solution that uses the file system as the workspace, chat as a temporary container, and persists research results in structured files.

File System as Workspace

Each research problem corresponds to an independent folder, storing structured files such as goals, sources, and claims. Its advantages include persistence, auditability, collaboration, and recoverability.

Source Quality Grading System

Establish a three-level source standard (A/B/C): Level A is original sources (highest credibility), Level B is strong secondary sources (aiding understanding), Level C is discoveries/noise (only guiding discoveries). Clear rules include: Level C cannot support final claims, and important claims must trace back to original sources.

Auto-Generated File Structure

research/my-research-topic/ contains files like GOAL.md (research goals), STATE.md (current state), SOURCES.md (source index), claims.md (claims and evidence), as well as directories like sources/ and notes/.

Collaboration Process with AI Agents

1. Read GOAL.md and STATE.md
2. Use references from SOURCES.md and claims.md
3. Decompose search sub-problems
4. Prioritize original sources
5. Save source texts
6. Update SOURCES.md and claims.md
7. Write synthesis content
8. Update STATE.md

Tool Agnosticism

Not tied to specific AI tools or search engines; can be used with Hermes Agent, Claude Code, OpenAI Codex CLI, etc., reflecting the lasting value of the research method.

Privacy and Security

The research/ directory is excluded from gitignore by default to avoid accidental publication of real research folders, protecting source confidentiality and data security.

Dimension	Traditional AI Chat Mode	Agentic Research Harness
Persistence	Lost after conversation ends	Files saved permanently
Traceability	Difficult to trace claim sources	Each claim is bound to evidence
Collaboration	Hard to share research state	Folders can be shared directly
Tool Dependency	Tied to specific AI platforms	Tool agnostic
Quality Control	Relies on AI self-restraint	Structured quality grading
Context Management	Limited by context window	File system has no limits

Agentic Research Harness represents a reflection on AI-assisted research, emphasizing that AI is a tool and chat is a means; the core value is verified knowledge and a clear thinking process. By persisting research states through the file system, the framework cultivates rigorous, transparent, and sustainable research habits, which are particularly important in the era of information explosion and AI hallucinations.

1. Create a research project
2. Fill in GOAL.md
3. Launch the AI agent from the project folder
4. Ask the agent to develop a small research plan
5. Search and save sources
6. Update SOURCES.md
7. Convert discoveries into claims.md entries
8. Write synthesis content to notes/ or outputs/
9. Update STATE.md
10. Start a new conversation when context is too large (no need to worry about loss as state is saved in files)