Research

Technical assets and open directions in agent reliability, shared memory and code intelligence.

Engineering Notes & Case Studies

Engineering Delivery & Evaluation for Agent Systems

Sep 2026

Transforming multi-step agent reasoning and tool calling into reproducible, verifiable, and auditable delivery loops across three layers.

From Lab Discussions to Pre-action Audits: Inside LabMemory

Sep 2026

Separating original, extracted and confirmed values, then using experiment results to update claim versions, evidence and scope.

Research & Technical Assets

Current work

Design-Agent Delivery & Evaluation

Building a reproducible acceptance path for PRD-driven design agents, from execution and native design output to structural and visual evidence; separating delivery validity, evidence completeness and requirement fidelity.

Design AgentFigmaEvaluationEvidenceReliability

Technical asset

Agent Shared Memory Bus

Local-first MCP memory infrastructure for Claude Code, Codex, Cursor and other clients to share persistent project memory. JSONL events, retrieval indexes, tiered lifecycles and cross-tool retrieval, with optional remote embedding backends.

MCPMemoryBM25SQLiteReliability

Active research

Cross-Agent Belief Divergence

Measuring cross-agent factual belief divergence as an early-warning signal for silent failure in multi-agent shared memory. τ/τ²-bench, environment oracle and pre-action audit.

Multi-AgentSilent FailureEvalResearch

Competition completed · Technical asset

LabMemory (晶研智流)

Trusted experiment memory built around claim, evidence, version, state, scope and responsibility, with pre-action audit. Annual Top 40 in the 2026 AI Pioneer Future Talent Competition, ending at the South China stage. Led the team and product / trusted-memory design; retained as an engineering case study.

Trusted MemoryPre-action AuditFeishuProduct

Active project

RepoMind Code Intelligence

Multilingual code knowledge base + MCP tool system with incremental code graph and trace observability for coding agents.

Code GraphMCPTraceRAG