Claude Code Opus 4.7/Sonnet 4.6 & token/memory crisis
Key Questions
What is the current status of the Claude Code token and memory crisis?
The crisis is at its climax, driven by context window resets, knowledge rot, and overloads in large codebases. Playbooks recommend CLAUDE.md files first, agentic search over RAG, and 3-6 month maintenance cycles using tools like GitHub Projects orchestrators.
How does Anthropic's Claude Opus 4.7 address context limits?
Claude Opus 4.7 is now generally available with a 1M context window, though users report it still suffers from forgetting and hallucinations beyond certain thresholds. Hybrid-reasoning capabilities help but require careful context engineering.
What are the weekly usage limit changes for Claude Code?
Weekly limits are increasing by 50% until July 13 to ease pressure on heavy users. This temporary boost aims to support expanded repo work and subagent experimentation.
What is Memory 101 for Claude Code agents?
Memory 101 explains that each session starts with a fresh context window that disappears when closed, with no built-in persistence. Techniques like δ-mem, Routines, and Manus Wiki help maintain state across interactions.
How does context engineering differ from traditional RAG?
Context engineering uses filtering, autonomous agents, and multi-model pipelines to package relevant information more effectively than basic retrieval. It addresses overloads better in large codebases by prioritizing hygiene and logging.
What tools help mitigate /goal failures in Claude sessions?
Fresh sessions combined with GitHub Projects as orchestrators and subagents are recommended to recover from failures. These approaches maintain continuity without relying solely on raw context size.
Why is MCP hygiene important in Claude workflows?
MCP hygiene, along with RAG, logging, DCQ, and Grafana observability, prevents resets and rot by standardizing memory and context management. It supports scalable agentic development in expanding repositories.
What expansions are planned for Claude Code repositories?
Repo expansions include subagents, Routines, δ-mem, Manus Wiki, and OpenClaw to boost performance by 50%. These build on new memory techniques and context engineering practices.
Crisis peaks with playbook (CLAUDE.md first, agentic search > RAG, 3-6mo maintenance), /goal failures prompting fresh-session + GitHub Projects orchestrators; 50% boosts till Jul13, repo expansions/subagents/Routines/δ-mem/Manus Wiki/OpenClaw; new Memory 101, context engineering, MCP hygiene, RAG/logging/DCQ/Grafana o11y/Prismatic Skills address resets/rot/overloads.