Claude Code Cache Bug Breaks Session Resume

A developer working on a multi-file refactoring task discovered their Claude conversation had lost all context after a brief interruption. The AI assistant, which moments earlier had been tracking changes across a dozen Python modules, suddenly couldn’t recall any of the previous discussion. This wasn’t an isolated incident—a caching bug affecting Claude’s prompt caching feature has been disrupting development workflows across Anthropic’s platform.

Background on the Caching System

Claude’s prompt caching feature, introduced in August 2024, allows the model to store and reuse portions of conversation context between API calls. The system caches up to the first 32,768 tokens of a prompt, reducing both latency and costs for applications that repeatedly reference the same information. Documentation at https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching explains that cached content remains available for five minutes of inactivity.

The feature works by identifying static content—like codebases, documentation, or system instructions—and storing it server-side. Subsequent requests that include identical cached content skip reprocessing, cutting response times by up to 85% according to Anthropic’s benchmarks. Developers quickly adopted caching for code review sessions, long-running debugging conversations, and applications requiring extensive context windows.

Key Details of the Session Resume Failure

The bug manifests when users attempt to resume conversations after brief pauses. Despite the five-minute cache window, sessions lose their entire context prematurely. Users report the issue occurring across different interfaces, including the Claude web application and API implementations. The problem appears most severe in coding scenarios where developers maintain extended context about project structure, variable naming conventions, and architectural decisions.

Technical investigations revealed inconsistent cache invalidation behavior. Some sessions maintain context for the full five minutes, while others drop cached content within seconds. The inconsistency suggests a race condition or distributed system synchronization issue rather than a simple timeout misconfiguration. One developer documented the problem with this API call pattern:

import anthropic

client = anthropic.Anthropic(api_key="your-key")

# Initial request with cacheable content
response1 = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=1024,
    system=[
        {
            "type": "text",
            "text": "You are analyzing this codebase...",
            "cache_control": {"type": "ephemeral"}
        }
    ],
    messages=[{"role": "user", "content": "Review main.py"}]
)

# Follow-up 30 seconds later - cache should hit
response2 = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=1024,
    system=[
        {
            "type": "text", 
            "text": "You are analyzing this codebase...",
            "cache_control": {"type": "ephemeral"}
        }
    ],
    messages=[{"role": "user", "content": "Now check utils.py"}]
)
# Cache miss occurs despite identical system prompt

Developer Reactions and Workarounds

The developer community responded with frustration on platforms like Reddit and Hacker News. Many reported abandoning multi-session workflows entirely, instead cramming all questions into single extended conversations. Others implemented manual context reinjection, copying previous responses into new prompts—an approach that defeats the cost and latency benefits caching was designed to provide.

Some teams built monitoring systems to detect cache failures. These implementations track cache hit rates through Anthropic’s usage headers and automatically retry requests when unexpected cache misses occur. While functional, such workarounds add complexity and don’t address the underlying reliability issue.

Broader Impact on Development Workflows

The caching bug undermines trust in Claude’s reliability for production applications. Development teams building AI-assisted coding tools had designed architectures around the assumption of stable five-minute cache windows. Applications for code review automation, documentation generation, and interactive debugging all depend on consistent context retention.

The issue highlights broader challenges in distributed AI infrastructure. Unlike traditional caching systems where failures typically result in performance degradation, LLM cache failures cause functional breakdowns—the model genuinely cannot continue conversations without context. This creates a binary success-or-failure scenario that’s particularly disruptive for interactive use cases.

Anthropic has acknowledged the reports but hasn’t published a timeline for resolution. The incident serves as a reminder that even sophisticated AI features require the same operational rigor as traditional distributed systems. For developers, the current recommendation remains conservative: design applications to gracefully handle cache misses and avoid critical dependencies on caching behavior until stability improves.