An always-fresh memory that learns your repo, so agents stop re-reading

Hi HN. Live-Memory is an open-source Claude Code plugin / MCP server that serves as an always up-to-date memory of your codebase. It uses a separate, low-cost, large-context-window model to distill your repo's organization, conventions and general structure. It observes the primary agent's activity and builds an understanding of your repo passively over time. This way, when you start a new Claude Code session, your primary agent queries Live-Memory via one read-only ask_live_memory tool and bootstraps its understanding of the codebase for the particular task at hand, instead of re-reading files. Live-Memory is shared among all Claude Code sessions (singleton). The memory model is pluggable and zero-config on a Claude subscription (Haiku), or point it at DeepSeek / any OpenAI-compatible / local model. Note it's a long-running local HTTP server you start once (not stdio).Benchmark results: in an A/B on a complex repo, on understanding-heavy work it offloaded ~93% of the building (premium) model's codebase-reading tokens, cut that model's cost ~61% per task and ran ~22% faster (edit-heavy work is break-even). Counting the memory model's own cost, it's ~25% cheaper all-in on Haiku — or ~57% on a cheap model like DeepSeek-v4-flash (which matched Haiku's accuracy at ~8x lower price), near the full 61% on a local model.Repo: https://github.com/shofer-dev/claude-code-live-memory — happy to go deep on the per-workspace memory window and passive ingestion.