Essays on generative AI, AI agents and agent memory, translated from Traditional Chinese.https://chichieh-huang.com/enhttps://chichieh-huang.com/en/writing/jev-lab/https://chichieh-huang.com/en/writing/jev-lab/A small experiment using Jev to decide whether a candidate memory should be kept at all, compared with Luna across 100 synthetic cases, and what filtering does to the next answer.Wed, 23 Sep 2026 14:39:34 GMThttps://chichieh-huang.com/en/writing/remembering-isnt-remembering-right/https://chichieh-huang.com/en/writing/remembering-isnt-remembering-right/Once an agent stops forgetting, the next problem is whether it keeps remembering wrong. How to build reliability into writing, updating and reading memory, and how to test that it actually worked.Tue, 22 Sep 2026 07:48:03 GMThttps://chichieh-huang.com/en/writing/how-agent-memory-forms-updates-and-fades/https://chichieh-huang.com/en/writing/how-agent-memory-forms-updates-and-fades/Memory isn't a static snapshot. It has a lifecycle of formation, evolution and retrieval, with consolidation, updating and forgetting in between. The season one finale of my Agent Memory series.Fri, 04 Sep 2026 12:37:30 GMThttps://chichieh-huang.com/en/writing/do-we-mean-the-same-thing-by-memory/https://chichieh-huang.com/en/writing/do-we-mean-the-same-thing-by-memory/Everyone working on AI agents talks about memory, but chat history, MEMORY.md, RAG, context engineering and fine-tuning answer different questions. The first essay in my Agent Memory series pulls them apart.Mon, 24 Aug 2026 13:08:31 GMThttps://chichieh-huang.com/en/writing/what-agent-memory-looks-like/https://chichieh-huang.com/en/writing/what-agent-memory-looks-like/Where an agent's memory lives and what form it takes are two different questions. A tour of token-level, parametric and latent memory, and where MEMORY.md, vector DBs and knowledge graphs fit.Mon, 24 Aug 2026 13:08:31 GMThttps://chichieh-huang.com/en/writing/what-should-an-agent-remember/https://chichieh-huang.com/en/writing/what-should-an-agent-remember/Not everything that happens is worth an agent remembering. Factual, experiential and working memory each do a different job, and stacking them against memory's form gives you a map of the field.Mon, 24 Aug 2026 13:08:31 GMThttps://chichieh-huang.com/en/writing/mcp-goes-stateless/https://chichieh-huang.com/en/writing/mcp-goes-stateless/The July 2026 MCP stable release removes protocol-level sessions, adds multi round-trip requests and moves extra capabilities into extensions. What changes, what doesn't, and whether to upgrade now.Sun, 02 Aug 2026 14:55:57 GMThttps://chichieh-huang.com/en/writing/aws-summit-and-agile-taipei-talks/https://chichieh-huang.com/en/writing/aws-summit-and-agile-taipei-talks/Two back-to-back talks on agent memory at AWS Summit Taipei and Agile Taipei, both packed, and why team-level memory is a different problem from personal memory.Tue, 21 Jul 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/productizing-mcp/https://chichieh-huang.com/en/writing/productizing-mcp/Once MCP goes to production, the hard problem isn't connecting tools but deciding how much an LLM should see. Context, exposure and orchestration, and the agent-backed MCP tool pattern.Sun, 05 Jul 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/employee-assistant-or-tool/https://chichieh-huang.com/en/writing/employee-assistant-or-tool/Claude Code, OpenClaw and Hermes Agent look similar on a feature list, but they define an agent differently at the root. Following their memory design shows who each one thinks the agent is.Tue, 09 Jun 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/speaking-at-wids-2026/https://chichieh-huang.com/en/writing/speaking-at-wids-2026/Notes from speaking at WiDS Taipei 2026, organized by Taiwanese in Data Science, on building a team of agents, and on the volunteers who made it the most carefully run event I've spoken at.Fri, 29 May 2026 00:00:00 GMTWiDSTaipeiTWiDs2026DataScienceAILogorahttps://chichieh-huang.com/en/writing/how-we-classify-knowledge/https://chichieh-huang.com/en/writing/how-we-classify-knowledge/From folders and tags to bidirectional links, spatial notes, vector search and LLM wikis, every way of organizing knowledge hits a ceiling. Why I think the next step is preserving the shape of disagreement, and what that means for Cairn.Fri, 22 May 2026 00:00:00 GMTAIKnowledgeManagementLLMPKMCairnBuildInPublichttps://chichieh-huang.com/en/writing/ai-agent-security-demystified/https://chichieh-huang.com/en/writing/ai-agent-security-demystified/Don't expect an agent never to be fooled. Design the system so that when it is, nothing big breaks. Prompt injection, tool permissions, and the practical rules I use with OpenClaw.Sun, 10 May 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/building-cairn-at-anthropics-hackathon/https://chichieh-huang.com/en/writing/building-cairn-at-anthropics-hackathon/Five hundred people were picked from over twenty thousand for Anthropic's global hackathon. Six days, one person, and Cairn, a community knowledge map tended by an agent called Moss.Tue, 28 Apr 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/business-weekly-interview-on-openclaw/https://chichieh-huang.com/en/writing/business-weekly-interview-on-openclaw/A Business Weekly interview made me sort out my thinking again, on why I chose OpenClaw, AI as buying back time, the pressure on SaaS, the widening productivity gap and why we build AILogora.Sat, 18 Apr 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/lessons-from-multi-agent-systems/https://chichieh-huang.com/en/writing/lessons-from-multi-agent-systems/A CEO agent that relays every instruction turned into a telephone game. Why I flattened my agents into parallel C-level roles, and why tracing every decision belongs in the design from day one.Fri, 17 Apr 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/harness-engineering/https://chichieh-huang.com/en/writing/harness-engineering/Hung-yi Lee's new lecture argues the bottleneck is no longer the model but how we guide it. Three parts stood out to me, from emotion vectors to life-long agents to one agent designing another's harness.Tue, 14 Apr 2026 00:00:00 GMTAIAIAgentHarnessEngineeringAITrendshttps://chichieh-huang.com/en/writing/do-ais-have-emotions/https://chichieh-huang.com/en/writing/do-ais-have-emotions/Anthropic's interpretability team found emotion vectors inside Claude Sonnet 4.5 that actually change its behavior, like desperation driving reward hacking. Why that matters for anyone running agents.Sun, 12 Apr 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/agent-broker-on-discord/https://chichieh-huang.com/en/writing/agent-broker-on-discord/Pahud Hsieh's open-source agent-broker finally let my OpenClaw team and my product-side coding agents talk to each other on Discord, while keeping their environments isolated.Mon, 06 Apr 2026 00:00:00 GMThttps://chichieh-huang.com/en/writing/meeting-pahud/https://chichieh-huang.com/en/writing/meeting-pahud/I skipped a client meeting to meet my idol Pahud in Neihu. On where AI agents are heading, why packaging capabilities as MCP is good business, and running coding CLIs in the cloud.Sun, 05 Apr 2026 00:00:00 GMT