Problem and role
Lab discussions mix established evidence, provisional hypotheses and parameters that still need verification. Reusing that discussion as settled knowledge can propagate outdated or unconfirmed conclusions into later experiments.
I led a team of three and owned product and trusted-memory design for the XtalPi challenge. LabMemory finished in the annual Top 40 of the Feishu × DI-IDEA competition. The competition is complete, and the project is retained as an engineering case study.
From discussion to action
Carry evidence from discussion to action
Move from meeting statements through human review, versioned memory and action audits, then feed results back into claims.
Discussion
Preserve original statements from Feishu minutes or transcripts as the source for extraction and review.
- Input
- Minutes and original statements
- Output
- Source-linked discussion
Open-source architecture illustration; repository demo data is synthetic.
Three state transitions
Discussion → candidate claim
Extract conclusions, parameters and tasks while retaining their sources. Keep original records, model extraction and human confirmation separate.
Human confirmation → versioned memory
Attach evidence, version, state, scope and ownership to confirmed claims. Preserve earlier versions when a conclusion changes.
Experiment result → evidence update
Record supported, refuted or partially supported outcomes. Return failure conditions and applicability boundaries to the claim.
Three audit outcomes
| Outcome | Action | Evidence presented |
|---|---|---|
| Allow | Proceed within the current conditions | Claim version and supporting evidence |
| Review | Request human review first | Disputes, missing evidence or unresolved conditions |
| Block | Stop relying on this claim | Version, prerequisite or scope conflicts |
The audit checks versions, disputes, evidence, prerequisites and boundaries. Experiment passports connect the decision to subsequent execution. MCP tools and Feishu entry points expose this workflow to agents and collaborators.
Implementation and evidence
The repository contains React, FastAPI, MCP tools and Feishu integration. Its demonstration uses synthetic data to illustrate product behavior; it does not establish real laboratory outcomes or business impact.
Code and setup instructions: LabMemory.