May 31, 2026
AI memory? Receipts, please
Show HN: Komi-learn – continuous memory and self-improvement for coding agents
This coding helper says it remembers your habits, but commenters want receipts
TLDR: Komi-learn is a new tool that tries to make coding assistants remember your preferences and past fixes automatically across sessions. Commenters say the idea is useful, but the real debate is whether it actually beats simple notes and whether anyone can prove it works in the real world.
A new Show HN project, Komi-learn, is pitching a very relatable fantasy: what if your coding assistant actually remembered how you like to work instead of acting like it just woke up from a nap every session? The tool watches what happened in one coding session, saves the useful lessons in the background, and brings them back next time automatically. No extra commands, no messy notes, no “wait, didn’t I tell you this yesterday?” energy. It also has an optional community pool where people can share approved lessons with each other, which gives the whole thing a slightly chaotic “group project, but with robots” vibe.
But the comment section immediately hit the brakes. The strongest reaction wasn’t “wow,” it was basically “cool story, where’s the proof?” The top response zeroed in on the biggest sore spot in this whole space: lots of flashy promises, not enough hard evidence that these memory tools work better than just keeping a few organized Markdown text files. Ouch. The commenter argued that what’s really missing is proper testing over long coding sessions, not more grand claims about self-improving agents.
That set the mood fast: part excitement, part side-eye, part weary veteran energy from people who’ve seen too many “this changes everything” coding tools before. The unspoken meme hovering over the thread is simple: your AI says it learns, the community says “show your homework.”
Key Points
- •Komi-learn is an early tool for coding agents that automatically recalls and distills durable lessons across sessions for Claude Code and Codex.
- •The project provides installation via pip or source, plus commands for setup checks, updates, configuration, syncing, contribution review, forgetting learnings, and uninstalling.
- •Its workflow is described as recall at session start, transcript distillation after the session, curation over time, and optional sharing to a community pool.
- •The article says sensitive or low-value information such as secrets, machine-specific paths, and one-off failures is filtered before LLM processing.
- •An optional community pool uses a GitHub repository of signed Markdown files, with user-approved PRs and BLAKE3/Ed25519-based addressing and signing.