ReAct Pattern: How Reasoning, Action, and Context Fix AI Chatbot Memory

·2 views

A developer building a customer support bot discovered that their initial implementation failed to retain conversation context, causing irrelevant responses. This led them to explore the ReAct pattern, a design approach used in agentic AI systems to help models maintain state across interactions. The pattern is composed of three components — Reasoning, Action, and Context — which work together to evaluate input, generate responses, and update conversation history. The developer applied this framework using LangGraph and MCP to build a state-graph-driven bot that routes customer queries based on detected intent. The article serves as the first entry in a 30-day series exploring AI agent design patterns.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer Builds AI-Powered Memorial Site for Abandoned Side Projects

A developer named Varshith V Hegde has launched DevGraveyard, a platform that lets developers ceremonially "bury" their abandoned GitHub side projects. Users connect their GitHub account, select an unfinished repository, choose a cause of death from options like "Never Made it Past Localhost," and write a short epitaph. Google Gemini's AI then generates a dramatic eulogy referencing the project's real commit history, including peak activity streaks and final commit messages. The platform also features a Three.js-powered interactive 3D cemetery with tombstones, candles, and fireflies, where the community can leave tributes or vote to resurrect projects. The site was built using Next.js 14, Supabase, and React Three Fiber, and is currently live at devgraveyard.varshithvhegde.in.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

E2LLM Open-Sources SiFR Browser Format to Limit AI Agent Attack Surface

E2LLM has published SiFR (Salience-Indexed Flat Relations), its browser capture format specification, as an open-source MIT-licensed project on GitHub. Unlike most browser-focused AI tools that function as autonomous agents with broad permissions, E2LLM positions itself as a perception layer that gives language models structured awareness of web pages without independent planning or looping. The SiFR format ranks and selects page elements by salience rather than dumping the full DOM or relying on accessibility trees, producing a compact 5–15 KB summary optimized for language model consumption. The MCP server exposes nine tools split between five read-only perception tools and four interaction tools, with state-changing actions subject to user confirmation. The capture engine and hosted server remain proprietary, while the format specification, taxonomy, and interface are fully open.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Eight Years of Freelancing Revealed Why Per-Word Pricing Punishes Good Writers

A freelance writer with eight years of experience in content and writing work shares hard-won lessons about pricing mistakes made along the way. Early in their career, they charged by the word, which inadvertently penalized quality — tighter, more skillfully crafted pieces earned less than padded, lower-effort ones. A key turning point came when they switched from quoting inputs like word count to quoting a flat project fee for a finished, publish-ready deliverable. The writer found that the biggest barrier to raising rates was not client resistance but their own self-doubt, noting that clients largely accepted higher flat fees without pushback. Their central takeaway is that per-word pricing structurally undervalues skilled writers, and that pricing should reflect the outcome delivered rather than the effort or volume logged.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Heirloom AI Uses Google Gemini to Archive Family Recipes and Traditions

A developer has built Heirloom AI, a web application that uses Google's Gemini multimodal AI to preserve family recipes, skills, and cultural traditions. Users can upload handwritten recipe cards, cooking photos, voice recordings, or videos, which the app processes into structured archival entries. Each entry includes poetic memory cards, ingredient checklists with confidence levels, and step-by-step physical-cue guides, with unclear details flagged for family verification. The full-stack app is built on React and Express, stores data in localStorage, and supports Markdown exports. The project was submitted as part of a DEV Community weekend challenge focused on passion projects.

0 comments Read more at DEV Community