ResearchMemoryKit sits between informal notes and heavier infrastructure.
Agent memory databases focus on retrieval: storing memories, searching them, and injecting them into future prompts.
ResearchMemoryKit focuses on project continuity:
- What is the current state?
- Why did we choose this route?
- Which failures should not be repeated?
- What evidence supports the current claim?
- What must be updated before a task is done?
It is intentionally file-based and git-native. There is no server, database, vector store, or hidden memory layer.
When automated guardrails are useful, the optional rmk.json contract and
rmk check validate project structure, router reachability, declared gate
headings, and Current State health without adding retrieval infrastructure.
P0 does not judge evidence quality or claim truth.
Experiment trackers are strong when a project needs dashboards, metrics, artifact storage, and run comparison.
ResearchMemoryKit does not replace them. It captures information that trackers often miss:
- direction decisions;
- failed-route rationale;
- environment pitfalls;
- evidence boundaries;
- current-state recovery;
- human-agent handoff notes.
The registry template stores metadata only. Large artifacts stay outside git or in a dedicated artifact system.
Project management tools track tasks. Research projects often need something different: a way to preserve uncertainty, negative evidence, and changing interpretations.
ResearchMemoryKit treats failed attempts and stop conditions as first-class records.
Notes apps are flexible, but flexibility creates drift when agents and humans return weeks later.
ResearchMemoryKit adds lifecycle semantics:
- overwriteable current state;
- append-only decisions;
- append-only failures;
- conclusion-preserving experiment logs;
- completion gates.
Use ResearchMemoryKit with:
- MLflow, Weights & Biases, TensorBoard, or custom registries for numeric run tracking;
- GitHub Issues for public task tracking;
- Zotero or bibliography managers for citation libraries;
- cloud storage for large outputs.
ResearchMemoryKit should remain the small, readable continuity layer that explains how these pieces fit together.