Nekobun — AI-Augmented Ebook Reader

Overview

Nekobun began at a hackathon with a simple question: what would reading feel like if the book could talk back — thoughtfully, in character, and with memory of everything you had read together?

It grew into an AI-augmented ebook reader built around a small cast of AI reading companions. Instead of one generic chatbot bolted onto a PDF, each companion had its own voice, reading style, and point of view — a skimmer who raced you to the gist, a close-reader who lingered on detail, and others in between.

The reading companions

The personas were the heart of the product. They are also where the hard problems lived.

For a companion to feel real across a 300-page book — and across sessions days apart — it needs two things most LLM apps don't have: a stable personality and durable memory of what you have read and discussed so far.

Why it was hard — and what it taught me

Two problems kept surfacing:

1. Long-term agent memory. Context windows fill up fast when a companion is meant to remember an entire book and your conversations about it. Naive history-stuffing broke down quickly.
2. Persona consistency. Keeping a character in character over long interactions — without drift, contradiction, or forgetting its own backstory — turned out to be an unsolved infrastructure problem, not a prompt-engineering one.

I couldn't solve these cleanly inside a consumer reading app, and I made the call to stop. But the problems were the real find: they are precisely what I went on to tackle in MemFuse, the open-source memory layer for LLM agents.

Status

Nekobun was archived in early 2025. The ideas didn't die — they moved down the stack. Nekobun was the product-discovery path that exposed a real infrastructure gap, and that gap became the next thing I built.