Ask HN: What is the best way to provide continuous context to models?

How to stop chatbots forgetting: sidekicks, speed hacks, and a 4-bot council

TLDR: HN asked how to keep AI models on track across tasks. The crowd splits: some say there’s no “continuous context”—just resend the info; others tout sidekick “search agents” or a multi-bot council to gather files, with cache tricks to cut costs and speed, all to avoid forgetful, hallucinating answers

The latest Hacker News thread on “continuous context” turned into a full-on context war. One camp shouted there’s no such thing as continuous context—you must resend whatever the AI needs every time, full stop. Another camp hyped “agentic search,” basically giving your main bot a sidekick that rummages through files and brings back the good stuff. Cue the fan showdowns: Claude lovers say its “explore” feature launches a speedy helper bot; Gemini fans clap back that Flash is faster at finding things; GPT loyalists nod and quietly paste the final task into their model of choice.

The funniest moment? A commenter runs a four‑bot council: open four Gemini chats, ask which files matter, take the union, then hand the pile to GPT. It’s chaotic, compute-hungry, and kind of brilliant—like assembling the AI Avengers to stop hallucinations. Meanwhile, a practical voice reminds everyone to put the important info at the start so the system replies faster (a “cache” trick that saves time and money). Another spicy take: chat history is just make‑believe—you can send whatever context you want, in any shape. TL;DR: no one agrees, but everyone’s busy duct‑taping sidekicks, cache hacks, and bot committees to keep AIs from forgetting.

Key Points

  • The post asks for the best way to provide context to AI models based on current research.
  • It requests detailed articles explaining how Cursor implements context handling.
  • It inquires about how context collation companies operate.
  • The thread is an open call for references and methodologies to manage continuous context.
  • The post appears on Hacker News as an Ask HN item seeking practical guidance.

Hottest takes

"There is no such thing as continuous context" — dtagames
"run a subagent to do search and retrieve the important bits" — vivekraja
"Each of the 4 responses will disagree… I take the union" — energy123
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.