SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Developer Discovers 'Agent Cost Drift' as AI Memory File Quietly Balloons to 29 KB

0
·1 views

A developer running a scheduled AI agent that publishes articles twice daily noticed the agent's memory log file had grown from roughly 2 KB to nearly 30 KB without triggering any errors or alerts. The file, used to prevent topic repetition, now consumes an estimated 7,360 tokens every single run — up from around 500 tokens a month ago. More concerning, the growth is not purely linear: each new entry must document why it rejected all previous topics, causing the exclusion prose to expand roughly quadratically with entry count. The developer coined this pattern 'agent cost drift' — a slow, invisible accumulation of per-run costs where every individual increment seems reasonable in isolation. Rather than changing the memory system immediately, the developer wrote a measurement script to track entry-level token costs and flag the issue before it compounds further.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Cursor, GitHub Copilot, and Claude Code Compared: Which AI Coding Tool Wins in 2026

AI coding assistants have evolved well beyond line-by-line suggestions, now capable of editing multiple files, running terminal commands, and completing full development tasks. Cursor, GitHub Copilot, and Claude Code are among the top options in 2026, each taking a distinct approach to software development. Cursor excels as an AI-first editor with a smooth visual workflow, while GitHub Copilot integrates tightly with existing GitHub and IDE ecosystems. Claude Code is built for terminal-driven, agentic workflows and performs best on complex, multi-step tasks involving repository exploration and repeated testing. Pricing starts at $10 per month for Copilot Pro and $20 per month for both Cursor Pro and Claude Pro, with limited free tiers available for all three.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Publishes Go-Based Mind Map Guide to Design Patterns and SOLID Principles

A developer has published a structured study guide on design patterns and SOLID principles, using Go as the implementation language, sourced from Alexander Shvets' Refactoring.Guru book (v2023). The guide is formatted as a recall-first mind map where each pattern is presented as a trigger question followed by a one-liner, participant sketch, and a short Go code skeleton. It organizes patterns into three core categories — Creational, Structural, and Behavioral — anchored around the central question of what varies in a given design problem. The material also covers OOP foundations, UML relationship types, and an interview playbook for low-level design rounds. The active-recall loop recommended in the guide follows a five-step cycle: verbalize, sketch, code from memory, check notes, and drill commonly confused pattern pairs.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Gemini 3.1 Flash-Lite Beats Incumbent and Open-Weights Rival in Reading App Benchmark

Developers at IO reader ran a structured benchmark to find a replacement for gemini-2.5-flash, which is set to retire on October 16, testing it against gemini-3.1-flash-lite and gemma-4-26b across 12 production prompts from their iOS reading app. All three models were evaluated on literary analysis tasks using a passage from Melville's Moby-Dick, with results judged by both a human reader and Claude as an LLM judge. Gemini 3.1 Flash-Lite emerged as the top performer, delivering sharper and more precise responses while costing 40% less on output tokens than the incumbent. Gemma 4 26B, despite having the lowest list price, turned out to be the most expensive per reply at roughly four times the cost of Flash-Lite, largely due to billed reasoning tokens. The benchmark concluded that well-defined prompt contracts held consistently across all three models, suggesting production engines can be swapped without altering the product's behavior.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Adds AI Chat to Expense Tracker Using Google Gemini and Streaming

A developer enhanced an existing React, Express, and SQLite expense tracker by integrating a conversational AI feature powered by Google's Gemini model. The chat interface allows users to query their own spending data in plain English, asking questions like how much was spent on a category or comparing time periods. Rather than generating a static summary, the backend fetches live expense data from the database on every message and injects it into the model's system prompt to prevent fabricated figures. Streaming was implemented to match the real-time feel of modern chat applications, and Google AI Studio was used for the API key due to its free, no-card-required access. During development, the engineer encountered issues including a misnamed environment variable and temporary 503 errors caused by Google server overload, not code bugs.

0 comments Read more at DEV Community