Tool slashes token use 95%

A free Claude Code plugin called claude-mem says it can cut token use by 95% by saving and reusing project context across sessions. (x.com) The tool is distributed as a Claude Code plugin from developer Alex Newman, who uses the handle thedotmack, and its public site showed 46,439 GitHub stars when checked on April 14, 2026. (github.com) (claude-mem.ai) Claude Code sessions normally start with a fresh context window, which means the assistant does not automatically remember prior chats unless developers load instructions or saved notes. Anthropic’s documentation says persistent context in Claude Code comes from CLAUDE.md files and an “auto memory” system that loads notes at the start of each session. (code.claude.com) claude-mem tries to solve that by acting like a running notebook for the coding agent. Its GitHub description says it captures what Claude does during coding sessions, compresses those observations with artificial intelligence, and injects relevant context back into future sessions. (github.com) (claude-mem.ai) The token claim is about how much text a model has to reread. Instead of resending large chunks of code and prior discussion every time, the plugin says it starts sessions with a lightweight index and fetches fuller records only when needed. (claude-mem.ai) Anthropic has separately pushed token-saving features in its own platform. In a March 13, 2025 post, the company said prompt caching on the Anthropic Application Programming Interface could reduce costs by up to 90% for long prompts, alongside cache-aware rate limits and more token-efficient tool use. (claude.com) The latest public claude-mem release was version 12.1.0, posted about four days before April 14, 2026, and the release notes describe “Knowledge Agents” that build searchable corpora from prior observations. The project’s release page also showed more than 48,000 GitHub stars, higher than the 46,000 figure cited in the social post. (github.com) That gap points to the main caveat in the viral claim: the adoption number is easy to verify only indirectly through GitHub stars, not through paid seats or active daily users. The plugin is plainly attracting attention, but the 95% savings figure appears to come from the tool’s own marketing and user demonstrations rather than an Anthropic benchmark. (x.com) (github.com) (claude-mem.ai) The pitch is simple: make a stateless coding assistant behave more like a teammate with notes. Whether the promised savings hold in practice will depend on how much repeated context a developer was sending into Claude Code in the first place. (code.claude.com) (claude-mem.ai)