Claude-Stream-Compactor
Smart hysteresis compaction for Claude Code, but limited to single API and not a general problem.

One tagline and a screenshot with no working demo or docs.
Developers working with LLM context limits
Smart hysteresis compaction for Claude Code, but limited to single API and not a general problem.
Cross-platform UI protocol with 97% token compression vs JSON—MCP for desktop use.
Five integration methods (MCP, CLI, API) beat single-method agent memory alternatives.
The core idea — turning agent-run debugging sessions into a reusable, searchable corpus (symptom + logs + minimal repro + env + stepwise fixes) — is smart and directly tackles an annoying repetition in agent workflows. The author even reports concrete time savings in a small benchmark, and the curl-first requirement (serve raw .md) is a blunt but effective attempt to avoid summarization loss. Big questions remain around verification signals and resistance to prompt-injection / brigading, so the concept is useful for people building agent infrastructure but not yet a broadly compelling platform.
Persistent memory for Claude Code with auto-compaction recovery—but relies entirely on external API.
15-18pp fact recall boost over flat summarization with full provenance tracing.