Claude Code's 25.5k token fixed cost — kou trims Memory to 4.2k

The article is a first-person measurement exercise, not a vendor announcement. kou starts from a familiar frustration — a 200k-token conversation window filling in a few turns and a weekly quota nearly exhausted by midweek — and decides to map where tokens actually go before paying for a higher plan.

The sequence matters. The first stop is the cost that exists before any conversation: running /context in a fresh Claude Code session reveals System prompt, System tools, Skills, Memory and Custom agents already loaded. Only then does the piece move to caches, long conversations, reasoning levels and code-search tools, and each later finding is framed as a layer on top of that fixed base. Memory is where kou found the clearest gain, cutting 7.3k to 4.2k tokens by deleting expired, duplicated and unused entries, and adding a roughly 60-token, two-line rule to CLAUDE.md to keep it from creeping back.

The stakes here are modest but concrete: whether the reader can keep working on a $20 monthly subscription rather than upgrading. That hinges on whether the same pattern — a large fixed load, stale memory, repeated prefixes — holds in their own repository, since the boundaries kou reports, such as the felt change around 350k tokens, are described as personal impressions that vary by task.