hSUM: Stateful Understanding & Memory
Local codebase evidence your coding agent can cite
Start new thread

hSUM: Stateful Understanding & Memory - Local codebase evidence your coding agent can cite

by•
hSUM gives coding agents cited answers from the repository they are working in. It builds a local, project-scoped lexical index and returns exact or BM25-ranked passages with stable hsum:// citations through MCP. No embeddings, hosted corpus, account, model, daemon, or telemetry. Install it once, authorize repositories as you work, and let agents call four read-only evidence tools. It works offline after the build and coexists with claude-mem and company-brain tools.

Add a comment

Replies

Best
Maker
📌
Hey Product Hunt 👋 I built hSUM because coding agents are good at generating changes, but their claims about a repository can be difficult to verify. Context disappears, memory gets stale, and hosted retrieval often asks you to upload your source code. hSUM adds a small evidence layer beneath the agent. It builds a local index for each authorized repository, exposes four read-only MCP tools, and returns exact passages with stable hsum://v1 citations. Retrieval is lexical/BM25, not embedding-based. There is no model, account, daemon, telemetry, or corpus upload. The setup goal is simple: paste one installation prompt into your agent, approve the macOS + Linux alpha warning when required, and let it handle installation, MCP registration, and activation of the current repository. New repositories can be authorized as you work. hSUM can also run beside claude-mem or a company-brain system because they solve different problems. hSUM provides evidence from the current source tree; memory tools preserve past sessions, decisions, and shared knowledge. This is an alpha, and I’m especially looking for feedback on retrieval quality, multi-repository use, and anywhere the onboarding still creates friction. If you try it, I’d love to know which coding agent and repository shape you used.