Komi-learn: Continuous Memory and Self-Improvement for AI Coding Agents
A curious thing happened on Hacker News this weekend. Amid the usual flood of Show HNs, a small project called Komi-learn climbed the front page — and it wasn't flashy. No demos of AI-generated UIs. No autonomous browser agents. Just a single, brutally useful idea: what if your coding agent remembered how you work, session to session, without you having to tell it?
The project, by kurikomi-labs, is a Python library that plugs into Claude Code and Codex (and potentially other agentic coding tools). It watches your sessions, distills durable lessons in the background, and loads the relevant ones at the start of your next session. No slash commands, nothing to save by hand.
The Memory Problem in Coding Agents
Every developer who has used Claude Code, Codex, or similar coding agents has experienced the same frustration: you teach the agent something in one session — your project's coding style, your preferred testing framework, that clever workaround for a third-party API bug — and by the next session, it's as if the conversation never happened.
The agent starts fresh each time, zero context from previous interactions.
Existing solutions exist, of course. Claude Projects let you upload instructions, but those are static documents you must maintain manually. Custom instructions in various tools are limited in length and scope. Hermes Agent pioneered the idea of automatic lesson distillation, but it runs as a standalone agent platform rather than a plugin for existing coding agents.
Komi-learn takes a different approach: a plugin layer that sits on top of your existing coding agent and gives it memory.
How Komi-learn Works
The architecture is refreshingly simple, built around three distinct phases:
1. Recall — "Here's what we learned last time"
When you start a new session, Komi-learn scans your previous learnings and loads those relevant to the current context. If you're working on the same project as yesterday, it surfaces the style preferences, dependency gotchas, and architecture decisions you established then.
2. Distill — "Here's what you should remember"
After each session, a background pass reads the entire transcript and extracts durable lessons: corrections you made, techniques that worked, fixes for recurring issues. It filters out noise like:
- Secrets and credentials (blocked by deterministic checks before the LLM even sees them)
- Machine-specific paths that won't apply across environments
- One-off failures that aren't worth remembering
- Complaints about tools that don't represent useful knowledge
3. Curate — "Is this still relevant?"
Over time, Komi-learn merges overlapping lessons and archives stale ones. The memory stays lean and actionable rather than accumulating noise.
The Community Pool: Shared Agent Wisdom
The most interesting feature might be the optional community pool. If you opt in, general-purpose lessons — ones that aren't project-specific or identifying — can be contributed to a public GitHub repo of signed Markdown files (no server required).
Contributions are:
- Scrubbed of anything identifying
- Reviewed by you before they leave your machine (each one opens a PR)
- Content-addressed using BLAKE3 hashes
- Signed using Ed25519 cryptographic keys
- Ranked by how many distinct GitHub accounts have signed a given lesson
This creates a Sybil-resistant reputation signal for community knowledge. If 50 accounts have signed a lesson about "how to structure FastAPI projects for maximum Claude Code performance," that lesson gets pulled before one signed by a single account.
"The idea is from Hermes Agent; this is my own take, generalized across hosts with an optional shared layer."
— Komi-learn README
What Makes It Different
| Feature | Static Instructions | Hermes Agent | Komi-learn |
|---|---|---|---|
| Auto-extraction | ❌ Manual | ✅ Agent-native | ✅ Plugin-based |
| Works with Claude Code | ✅ Basic | ❌ Standalone | ✅ Native plugin |
| Works with Codex | ✅ Basic | ❌ | ✅ Native plugin |
| Community knowledge pool | ❌ | ❌ | ✅ GitHub-based |
| Cryptographic signing | ❌ | ❌ | ✅ Ed25519 |
| Privacy-first | N/A | ✅ | ✅ Scrubs + approval |
The key differentiator is optionality. Komi-learn doesn't force you to adopt a new agent platform. It layers onto the tools you already use, which dramatically lowers the switching cost.
Installation and Getting Started
Getting started takes minutes:
pip install komi-learn
komi-learn install
The install command runs an interactive setup, verifies connectivity with a real model call, and hooks into your agent's startup flow. If a model isn't reachable at runtime, it silently skips the learning pass rather than interrupting your session.
For the impatient, there's even a zero-setup demo:
python examples/demo_loop.py
This runs two sessions: you correct the agent in the first, and the second shows it recalling the correction without any additional input.
The Bigger Picture
Komi-learn's rise on Hacker News signals something important about where the coding agent ecosystem is heading. The first wave of coding agents was about capability — can they write code? The second wave is about continuity — can they build on what they've already done?
Projects like Hermes Agent pioneered the concept of automated lesson distillation. Komi-learn extends it by making the idea portable across agent platforms and adding a cryptographically signed community knowledge pool — a decentralized repository of agent best practices that grows organically.
This is the direction the industry is moving: agents that get better the more you use them, without requiring manual configuration or prompt engineering. The next frontier isn't better models — it's better memory architectures that let agents learn from experience.
(Sources: Komi-learn GitHub, Hacker News discussion, Hermes Agent)
Cet article a été initialement publié sur The Agent Report.
Top comments (0)