Technical Analysis

Why Claude Code Loses Context

Context loss, session amnesia, context bleeding — it's not a bug. It's an architectural constraint called context decay. And it's solvable.

The Core Problem: Context Loss & Session Amnesia

Claude Code operates within a 200K token context window. Every message, every file read, every error, every retry — it all accumulates in that window. There is no automatic garbage collection.

After approximately 60 minutes of active use, the context window begins filling with:

  • Failed attempts and error messages
  • Verbose file contents from every read operation
  • Correction history and patch-over-patch sequences
  • Stale conversation turns no longer relevant to the task

As the context fills, the original instructions get pushed further away from the model's attention. By hour 2, Claude literally cannot "see" what you told it at the beginning. This is why it loses track of your architecture, contradicts itself on design decisions, and exhibits what developers call "context drowning" — the agent is submerged in noise and can no longer surface the signal.

The Degradation Timeline

0-30 min
Recall: 95%+

All instructions fresh. High-quality output.

30-60 min
Recall: 80-90%

Subtle drift begins. May repeat suggestions or miss constraints.

60-120 min
Recall: 40-65%

Architectural decisions forgotten. Agent begins contradicting earlier work.

120+ min
Recall: <40%

Agent patching its own patches. Session effectively corrupted.

Why CLAUDE.md Alone Doesn't Fix This

The CLAUDE.md system prompt is loaded at session start, but it occupies tokens in the same context window. As the session grows, even the system prompt gets less attention weight.

"System prompts are natural language suggestions, not deterministic constraints. Under context pressure, agents routinely bypass text-based instructions."

The Governance Solution

Runtime governance prevents context rot through deterministic enforcement, not natural language suggestions:

  • 1. Checkpoint Rotation — automatically prune stale context at 65% utilization while preserving architectural state
  • 2. Semantic Reset — emergency context purge at 85% that reloads core instructions
  • 3. Patch Chain Detection — halt execution when an agent patches the same file 3+ times
  • 4. Session Duration Limits — mandatory checkpoints every 30 minutes
Deploy Context Rot Prevention →View Context Degradation Data

Frequently Asked Questions

Why does Claude lose context and get session amnesia?
Claude operates within a finite context window. As conversations grow, original instructions get pushed out of the attention window — causing context loss, session amnesia, and context bleeding between tasks. The agent loses track of architecture decisions and starts contradicting itself.
Does this happen with all AI coding agents?
Yes. Context decay affects Claude Code, Cursor, Windsurf, Cline, and every agent that operates within a finite context window. The rate varies by model, but the pattern — context drowning, instruction drift, and eventual session amnesia — is universal.
Can I just start a new session?
Starting a new session resets context but loses all accumulated understanding of your codebase. The real solution is checkpoint rotation — preserving architectural state while pruning conversation history to prevent context window bankruptcy.
How much does context rot actually cost?
In documented incidents, context rot has caused 6+ hours of rework, 23 corrupted files, and $340+ in wasted tokens in a single session. Across a team, this compounds into $100K+ annually.
← Return to Infrastructure Catalog