Why I Send the Full Conversation History on Every Claude Turn

Every developer I talk to truncates their conversation history to save tokens. I do the opposite — I send the full history every turn. Here's why, and why it actually costs less.

The truncation trap

When you truncate conversation history, you break Claude's cache prefix. Anthropic's prompt caching works on exact prefix matches — truncation at turn 15 means a different cache key than turn 14. You pay full price every time.

The math that surprised me

Full history: cache hit rate ~85%, effective cost $0.0003/1k tokens. Truncated history: cache hit rate ~12%, effective cost $0.0027/1k tokens. Sending more tokens costs less.

When to actually truncate

There are legitimate reasons to truncate — tool result floods, image-heavy sessions, very long document inserts. Here's the decision rule I use.

How to implement it cleanly

The key is separating static context (system prompt, user profile, tool definitions) from dynamic context (conversation turns). Cache the static layer aggressively, let the dynamic layer grow.