Stop context rot in long sessions
Keep Claude sharp across multi-hour work — prevent the slow degradation that happens when the context window fills up with irrelevant history.
Every long Claude session hits a point where the model starts forgetting earlier decisions, re-reads files it already has, or generates output that contradicts work from an hour ago. This is context rot — and it's silent. No error message tells you it's happening; the quality just drops. The fix is to deliberately checkpoint and isolate context instead of letting it grow until it breaks.
Recommended
The 3 tools we'd reach for first
Paste this into CLAUDE.md to force a structured checkpoint whenever context usage crosses 70% or 20 tool calls. The brief captures what's decided, what's open, and the exact next step — so a fresh session continues seamlessly.
Drop-in template for the TOP of a new session when continuing work. Cuts the 'warm-up tax' to near zero and prevents Claude from making contradictory assumptions about prior state.
When a task involves 5+ files or 10+ tool calls, delegate to a subagent with a fresh context window instead of accumulating everything in your main conversation. Pattern includes clear criteria for when to delegate.
Alternatives
Use these when your stack calls for them
Multi-file context organization pattern — `about-me.md`, `brand-voice.md`, etc. Better fit for teams with stable long-running projects where the context is shared.
Supporting
Helpful adjacent tools to complete the workflow
External memory engine if you need persistence beyond a single project. More moving parts than the paste-in patterns above.
Not sure where to start?
Browse all 409 tools in the full catalog, or explore by role or trust tier.
Browse catalog · By role · By trust tier