Building a memory architecture for GPT-6 Astra agents instead of piling on context
The author argues agents fail because developers treat the conversation as a database, and proposes compressing history, limiting tool calls and keeping only the goal, decisions and next action.
your GPT-6 Astra agent has a memory leak. you built it. most developers use the conversation as a database. every message, search result, failed tool call and status update gets carried into the next step. eventually Astra is not solving the task. it is searching through the wreckage of everything it already tried. the fix is not a larger context window. it is a memory architecture: previous_response_id continue from the last state context_management compress history that no longer changes the plan max_tool_calls stop failed tools from becoming infinite loops then save the information compaction must never destroy: • current goal • decisions already made • verified results • unresolved failures • next action the chat is temporary. the state is the product. a reliable agent doesn’t remember every word you sent it. it remembers what changed, what failed and what must happen next. I broke down the exact GPT-6 Astra memory architecture below ↓
This post quotes another post: x.com/mr_kozh/status/2100171343752798467