Agent Memory Is Not a Database: Start with an Event Log
I break agent memory into writing, storage, retrieval, assembly, and forgetting, then outline a minimal implementation path that starts with a durable event log.
Practical AI, explained.
I build practical AI systems, automate repetitive work, and write down what I learn.
Latest writing
Recent writing
View all writing →The Agent Extension Stack: Skills, Plugins, MCP, and Tools
A Skill teaches an agent how to work, a Tool lets it act, MCP connects it to external capabilities, and a Plugin packages the pieces for installation.
Reading World Models Through 'LLMs Can't Jump'
A reading note on abductive reasoning, interactive simulation, and what current world models can and cannot do.
What Pi Coding Agent Keeps in the Core
A source-based comparison of Pi, Codex, and OpenCode shows how a minimal coding harness leaves workflow policy and isolation to the user.
Testing ASR for Video Subtitles
A small Korean-language comparison showed why transcript coverage, timestamps, and manual review matter as much as model output.
Three paths through the work
Curated reading on systems that recur across the projects.
-
Agent Workflows I Can Actually Use
Field notes on using roles, tools, loops, and checks to make agent workflows useful.
-
Media Pipelines with Inspectable Stages
Field notes on separating transcription, translation, vision, and rendering into inspectable media stages.
-
Cloud Systems for Small, Real Work
Field notes on moving useful scripts and AI systems from one machine to dependable cloud services.
