addxai/reflectworld — explained in plain English
Analysis updated 2026-05-18
Give an AI agent long term memory of what happened in front of a camera over days or weeks.
Build an assistant that recognizes recurring people or objects and recalls facts about them.
Run the dashboard to upload video or connect a live stream and chat with the stored memory.
Study the accompanying research paper and benchmarks on multimodal memory systems.
| addxai/reflectworld | compartmentdev/compartment | denissergeevitch/pole-chudes-2 | |
|---|---|---|---|
| Stars | 77 | 77 | 77 |
| Language | TypeScript | TypeScript | TypeScript |
| Setup difficulty | moderate | moderate | easy |
| Complexity | 4/5 | 4/5 | 2/5 |
| Audience | developer | ops devops | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires Node.js 22 or newer plus a Python component for the perception pipeline.
ReflectWorld gives AI agents a persistent memory of the physical world captured through cameras. Instead of an AI only seeing short, isolated video clips and forgetting everything between them, ReflectWorld turns a live camera feed into an ongoing, searchable record of what happened, who or what appeared, what changed over time, and what an agent should remember later. The project describes itself not as a video summarizer but as a memory layer sitting underneath an AI agent. The README lists several core capabilities. It can take video from almost any source, including fixed cameras, RTSP streams, webcams, phone uploads, smart glasses footage, and robot cameras. Rather than treating every frame the same, a higher level scene agent decides what matters, keeping quiet moments lightweight while giving deeper analysis to important ones. Over time, it links related events into a continuous story rather than a pile of disconnected captions, for example turning several short moments of someone making coffee into one coherent scene. It also remembers recurring people and objects, building up sightings, roles, actions, and other facts about them each time they reappear, so a later moment can be understood in light of what the system already knows about that entity. All of this memory is made available to an AI agent as searchable evidence, meaning an agent can ask what happened, who was involved, or why something mattered, and get an answer grounded in stored moments rather than a guess based on a single clip. A bundled dashboard lets a person upload a video or connect a live stream, watch a timeline of moments fill in as analysis runs, inspect what entities and memories have been captured, and then ask questions in a chat interface that draws on that stored context. The README states that privacy is treated as a system boundary: raw source files, non language model recognition results, media, identity evidence, and the persistent memory itself stay on the user's own machine, and outside AI model providers only receive the specific inputs needed for whichever capability is explicitly enabled. The project cites an accompanying research paper and includes benchmark comparisons against other memory systems. Built primarily in TypeScript with a Python component, ReflectWorld requires Node.js version 22 or later and is released under the Apache 2.0 license.
A persistent visual memory system that lets AI agents remember people, places, and events from live camera feeds over time.
Mainly TypeScript. The stack also includes TypeScript, Python, Node.js.
Use freely for any purpose, including commercial use, as long as you keep the license notice and any stated attributions.
Setup difficulty is rated moderate, with roughly 1h+ to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.