AI agent forgets user preferences every session. This fixes it
I build AI agents for work and kept hitting the same issue: a user corrects the agent, the session ends, and the correction is gone. Next session, same correction. I tracked it across our users and the average preference gets re-corrected 4+ times before people just give up. Existing solutions don't really solve this. Memory layers store raw conversation logs. RAG retrieves documents. Neither extracts what the user actually wants as a structured, persistent preference. So I built pref0. It does one thing: extracts structured preferences from user corrections and compounds confidence across…
What it does
In the maker’s words, at launch
I build AI agents for work and kept hitting the same issue: a user corrects the agent, the session ends, and the correction is gone. Next session, same correction. I tracked it across our users and the average preference gets re-corrected 4+ times before people just give up. Existing solutions don't really solve this. Memory layers store raw conversation logs. RAG retrieves documents. Neither extracts what the user actually wants as a structured, persistent preference. So I built pref0. It does one thing: extracts structured preferences from user corrections and compounds confidence across sessions. How it works in practice. Say you're building a customer support agent: Session 1: User says "always escalate billing issues to a human, don't try to resolve them." pref0 extracts billing_issues: escalate_to_human, confidence 0.55. Session 4: User flags a billing ticket the agent tried to auto-resolve. pref0 reinforces the preference. Confidence hits 0.85. Session 7: A billing issue comes in. The agent routes it to a human without being told. No correction needed. Now multiply that across hundreds of users. Each one teaching your agent slightly different things. pref0 maintains a structured profile per user (or team, or org) that your agent reads before every response. The API is intentionally minimal. Two endpoints: POST /track: send conversation history after a session. pref0 extracts preferences automatically. GET /profiles/{user_id}: fetch learned preferences before the agent responds. A few design decisions: > Explicit corrections ("don't do X") score higher than implied preferences. Stronger signal. > Preferences are hierarchical: user > team > org. New team members inherit org conventions on day one. > Confidence decays over time so stale preferences don't stick forever. This isn't a replacement for memory. Memory stores what happened. pref0 learns what the user wants. You can run both side by side. Works with LangChain, CrewAI, Vercel AI SDK, or raw API calls. Free tier available -> https://pref0.com/docs Would love feedback on the approach, especially from anyone building agents with repeat users.
Does the same job
all alternatives →
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
