LLM Algorithm Engineer at Ant Group (阿福). I work on memory systems for LLM agents, self-evolving experiment loops, and LLM × graph computing. Before that: core maintainer at CAMEL-AI, GraphRAG maintainer at DB-GPT.
On memory
- Memory is not storage. It is external state a decision can use — the measure is whether the channel from history to the current decision works, not how much history is kept. → 几万字都讲不明白的 Memory 架构与思考
- Time is a first-class dimension. Bi-temporal, time-sliced recall turns "when was this true" into a hard constraint on retrieval and aggregation, not a metadata field. → same article
- Parametric vs. non-parametric memory is a question of where the write cost lands: compiled into weights at training time, or paid at commit and retrieve/inject time. The ceiling of the latter is interface bandwidth, retrieval-aggregation error and policy learning. → 参数化 Memory 漫谈
- Forgetting is something to design, not a bug. Optical compression (DeepSeek-OCR) hints at memory tiers where old context is re-rendered smaller and decays gracefully. → 用 8500 字解析 DeepSeek OCR 与记忆系统
On agents
- Fast and slow thinking as two models: a Thinker that plans and a cheaper Actor that executes tool calls — built in 2024, before "thinking" models existed. → Chat2Graph · Reasoner
- One active, many passive: a single Leader decomposes work into a sub-job DAG for many Experts, with recursion and error re-injection instead of a flat group chat. → Chat2Graph · Leader
- "Less structure": agent workflows should be searched and optimised — MCTS over declarative configs, context engineered per layer — not hand-written as SOPs. → Chat2Graph · Workflow generation, the OSPP 2025 project I mentored
- Research is itself an agent loop: falsifiable hypotheses → experiments → judge + verifier → attribution. Failed experiments are memory too — replay for the next hypothesis.
| Project | Role |
|---|---|
| Chat2Graph — graph-native agentic system | Lead contributor (#1 by commits and lines); OSPP 2025 mentor |
| Apache GeaFlow (incubating) — streaming graph engine | Contributor — CASTS, an LLM reasoning operator |
| DB-GPT — agentic AI data assistant | GraphRAG module maintainer, 2024 |
| CAMEL-AI — multi-agent framework | Core maintainer, 2023–2024; led the Mixture-of-Agents design |
| LeAgent | Author — an early (2023) RAG chatbot that cites its sources inline |
- 几万字都讲不明白的 Memory 架构与思考 — memory as ledger → views → policy, and time as a hard constraint. Also on AntData and OceanBase.
- 参数化 Memory 漫谈 — parametric vs. non-parametric memory (on 阿里技术).
- 用 8500 字解析 DeepSeek OCR 与记忆系统 — optical compression as a memory tier (on OceanBase).
- Chat2Graph design docs: overview · memory — DIKW layers · reasoner · workflow generation
- Alliance Sorbonne Université — engineering degree in computer systems, 2025
- Shanghai University — B.Eng. in information engineering, 2024
- Software engineering intern, Synopsys, 2023
- Ant Group AI X-STAR, 2026
- Ant Group Open Source Pioneer Award, 2024
- MCM/ICM 2022 Outstanding Winner (top 0.16%)
- Qian Weichang Presidential Scholarship — Shanghai University's highest undergraduate honour, 2022
- National Scholarship (国家奖学金), 2022
- Outstanding Graduate of Shanghai, 2024