Engineering research · Local repository tooling
Mímir
Stop rediscovering the codebase.
A coding agent can search a repository and still lack a useful understanding of it. The relevant function may be easy to find while the surrounding constraints, dependencies, and reasons for its design remain scattered.
Mímir applies the continuity problem to the codebase itself. The aim is a persistent repository model that can help an agent recover the right part of the system for the task in front of it.
That means connecting subsystems, files, symbols, boundaries, and invariants, then returning a bounded context bundle with its sources and uncertainty intact. The system also needs a way to widen its view—or admit that the available context is insufficient.
Mímir extends the memory work into repository understanding. Its engineering challenge is keeping that understanding useful as the code changes, while avoiding a flood of loosely related information.
A bounded view, with room to abstain
The implementation maintains local repository state, incrementally ingests changed files, and connects files, symbols, cards, and subsystem boundaries. Git history contributes file-level rename, move, and co-change context.
A deterministic broker returns bounded context with decision traces and explicit abstention. Remap and repair flows address stale bindings as the repository changes. Deeper semantic relationships and learned graph models remain development directions.
This page makes no measured token-saving or productivity claim. The useful distinction is ownership: Mímir models repository structure; Muninn preserves durable project and user memory.