Heimdall
A memory layer for AI coding agents that indexes work across every repository a developer touches into one semantic graph, so an agent can check whether a problem was already solved elsewhere.
Heimdall answers "did I already solve this in another project?" with one ranked search instead of a fresh round of grep, find and ls across repositories, and injects relevant prior work into a new session's first prompt. Indexing runs as a local daemon using file watching, tree-sitter parsing and locally computed embeddings, so it costs CPU time rather than LLM calls, with hybrid lexical, semantic and graph-walk retrieval at query time and optional GPU acceleration.
Retrieved results carry a verification status rather than being returned as unchecked matches: one status means the file is unchanged and the query's terms are actually present in the indexed text, another explains which check failed, and others mark a moved or dead file so the index keeps ranking accurately.
It is written in JavaScript and distributed as a CLI, licensed under MIT.