ExperienceOS

ExperienceOS gives AI agents a lifetime of experience by remembering, updating, forgetting, and recalling the right memories, enabling better decisions across sessions within limited context windows.

Comment

What it does

ExperienceOS is an experience layer for AI agents, powered by Qwen (qwen-plus) on Alibaba Cloud. It gives agents persistent memory across sessions by deciding what to remember, what to update when preferences change, what to forget when information becomes obsolete, and what to recall within a limited context window. Crucially, every one of those lifecycle changes is verified and authorized by an exact, single-use receipt before it can touch durable state, so the agent can update and forget aggressively without ever granting unrestricted mutation authority. As users interact, it makes increasingly accurate decisions without repeating the same information.

How I built it

I designed ExperienceOS as a provider-independent layer that wraps an LLM rather than replacing it. Qwen on Alibaba Cloud DashScope serves as the reasoning engine, while ExperienceOS manages the complete memory lifecycle —creation, retrieval, updates, forgetting, ranking, and context assembly— through deterministic controllers, a bounded runtime authorization step, and a single mutation boundary.

The live demo is deployed on Alibaba Cloud (an ECS instance running the app; DashScope serving inference), and every memory decision is exposed in the UI — a per-message receipt inspector plus an "engine room" trace, so users can watch experience accumulate, supersede, and be forgotten in real time.

Accomplishments that I'm proud of

I'm proud of building an experience layer rather than another chatbot and backing it with evidence. On a frozen lifecycle benchmark run live through Qwen, activating the memory lifecycle raised final-answer accuracy from 71% to 79%, cut stale-answer usage from 50% to 28%, and kept the prompt context roughly 87% smaller than replaying full history, comparable to the strongest memory baselines, with the four previously genuine stale-answer failures fixed. I also created a reusable architecture, with the memory mutation path guarded by exact authorization, that can attach to future agents instead of being tied to one app.

What's next for ExperienceOS

My next goal is to expand ExperienceOS into a complete experience platform for autonomous agents: additional model providers, richer memory types, collaborative experience shared across multiple agents, durable one-time authorization for stronger concurrency guarantees, and more intelligent memory policies, so agents keep improving throughout their lifetime while staying efficient within limited context windows.

Built With

Share this project:

Updates