Technology

When waiting, polling, and switching between Agent sessions keep breaking our focus, task-driven work offers another way: let machines execute asynchronously while humans think one thing at a time.

Notes from a real refactoring project on GitHub Stacked PRs: linear dependencies, layered merging, multi-level synchronization, and local branch relationships.

From Codex's hard removal of chat/completions to the migration damage it caused in new-api, this examines the real tool-flow differences between Anthropic Messages and OpenAI protocols—why my long-chain silence is not the protocol's fault, and what truly counts as talking while working.

A breakdown of Claude Desktop's five scheduling systems, the architectural causes behind their complexity, why this design multiplies the effort required to maintain one project, and what a bad architecture can teach us.

A discussion inspired by MoeChat about RRF and Hybrid Search: how should time retrieval participate in hybrid retrieval, how should it adapt across scenarios, and where are the boundaries between wikimem and XnneHangLab?

An in-depth look at how MoeChat builds long-term memory for AI characters, including its time-and-semantic hybrid retrieval—written for people interested in memory systems.

Starting from the question of what sparse and dense really mean, this works through the embedding pipeline, subwords and vocabularies, input length, and their practical uses.

My plain-language explanation of TF-IDF, from terms and tokenization through TF, IDF, query scoring, BM25, and using it to recommend related posts on this blog. Apparently, it is much easier to pay back the debt of skipping NLP lectures once there is a reason to learn it.

Does your Agent keep repeating the same mistake before it finally remembers? A look at gotcha.md: which memories must be fully available as soon as a new session begins, and which can wait to be retrieved later?

Reading notes on How We Use Skills: when you cannot put everything in front of an Agent at once, how do you keep it effective? Progressive Disclosure and Task Decomposition.

memU ADR 0007 overturns the earlier LLM mode: thoughts from a consumer-product developer and emotional-companion perspective.

A breakdown of memU focused on its main paths and ADR architecture documents, currently following through ADR 0008.

Drawing on our experience building an AI companionship engine, this article compares the limitations of mainstream Agent frameworks such as LangChain, Dify, LangGraph, and Coze, and documents the core design ideas and trade-offs behind the low-level framework we built for real-time companionship.

A hands-on introduction to using the Obsidian-YOLO plugin and why it is worth recommending.

Starting from hands-on work on an AI desktop companion's long-term memory, this post compares three long-term-memory approaches for LLMs and asks whether RAG belongs in a personal blog's relationship graph.

An investigation into why web and API results diverge during multi-image analysis, from Lost in the Middle to a two-stage approach based on per-image pre-summaries.

A Termix deployment guide: use Docker to host a web-based SSH client and finally say goodbye to Termius acting up.

LLM attention is limited. Starting from the difference in weighting between system and user prompts, this article explores the Lost in the Middle phenomenon and what it can teach us.

A guide to using OBS together with VTube Studio for livestreaming or video recording.