Most AI coding systems still behave like first-time visitors in unfamiliar repositories. Mímir maintains repository world-state and brokers bounded rehydration through a deterministic interface so the model can ask where to start, what is adjacent, and when to stop widening.
Featured diagram
Search-style retrieval versus structure-aware navigation inside a real repository.
Repository world-state
Brokered orientation
Bounded rehydration
The Problem
AI coding tools keep re-solving orientation from scratch.
The deeper failure is not bad code in isolation. It is the appearance of understanding before the system is actually oriented. Search, retrieval, and larger context windows still begin from the same disadvantage: reconstructing the repository from fragments instead of starting from a durable sense of what owns the behavior and what boundaries sit nearby.
Failure symptoms
fragmented reconstruction
shallow retrieval
mis-scoped edits
noisy widening
repeated reorientation cost
Where the stack bends
Traditional RAG pushes orientation work into retrieval and the context window instead of solving it structurally first.
What Mímir Changes
Mímir is repository orientation, not a search feature.
Mímir is a persistent repository self-model. It maintains world-state and brokers bounded context from likely subsystem to likely file to likely symbol. The shift is from "find similar text" to "start from an oriented map of the codebase."
Navigation model
Rehydration narrows from likely subsystem to likely file to likely symbol, then returns a bounded bundle with spans, summaries, adjacencies, confidence, and continuity where it matters.
Mímir is a GPS for a codebase, not a search engine.
Cards
Units of orientation about the repository: subsystem, file, symbol, boundary, workflow, or invariant.
Bindings
Links that keep orientation stable as code moves, is renamed, or is partially rewritten.
Signals
Evidence that shapes confidence through structure, traces, edit history, retrieval success, and verified outcomes.
Adjacencies
Relationships that reveal what is nearby, upstream, downstream, dependent, or risky to touch in the system.
How It Works In Practice
Orientation is delivered through a brokered boundary.
Mímir does not just store structure; it serves it through a brokered boundary between the model and world-state services. The model asks in explicit modes and receives either a bounded rehydration bundle with sources, adjacencies, and confidence or a disciplined abstention when widening would add noise. The boundary is deterministic and inspectable, creating a measurable eval surface instead of opaque prompt stuffing.
Brokered request modes
Locate: find the most likely place to begin.
Rehydrate: assemble the bounded working set.
Neighborhood: widen from confirmed anchors only.
Boundary: confirm ownership and safe edges.
What the model receives
Bounded rehydration bundles, not raw search hits.
Sources, adjacencies, and confidence signals.
Inspectable traces that define the eval surface.
Explicit abstention when widening would add noise.
Unit of delivery: a bounded rehydration bundle.
Current State
Operational first version, hardening underway.
Mímir has moved beyond thesis stage: the world-state substrate is implemented, the brokered interface layer is live, and the system has been exercised across real repositories. In those runs, brokered orientation usually finds the right place to begin; most misses now come from widening and abstention policy.
In place now
repository world-state substrate implemented and maintained
deterministic broker/interface layer between model and world-state services
bounded rehydration bundles delivered with sources, adjacencies, and confidence
proving across real repositories, not a single favorable sandbox
structured traces and policy tightening reducing supplement noise
Hardening focus
remaining work is precision, policy, throughput, hardening, and generalization
core weakness now is disciplined widening and abstention, not initial orientation
supplement policy decides when to widen, when to stop, when to defer
bounded temporal signal remains experimental rather than foundational
designed for messy mixed trees, not handpicked clean demos
Focus now: policy-governed widening, disciplined abstention, throughput, and generalization.
Why This Matters
Repository-scale reliability needs orientation, not just recall.
This is not just better search. It is an operating model for large, stale, unfamiliar, recently changed, or messy repositories: bounded working sets, better first-target selection, disciplined widening, and honest abstention when extra context would make outcomes worse.
Lower token waste
Less context budget spent reconstructing orientation before useful work begins.
Better first-target selection
Agents get pointed at the right subsystem, file, and symbol earlier in the task.
Higher trust in repo-scale work
Bounded working sets and honest abstention reduce failures caused by noisy widening.
Who This Is For
The people who can place and scale this layer.
Technical cofounders / operators
Builders who want to take orientation infrastructure from sharp system to durable product.
Developer tools / code intelligence
Leaders working on code navigation, search, and agent surfaces who see orientation as the missing layer.
Retrieval and agent infrastructure
People who can harden brokered context into a reliable service boundary.
Commercialization and category builders
Partners who understand how to position a foundational layer inside the AI coding stack.
Closing
If you're building the stack around AI coding systems, let's talk.
I'm looking for technical, product, and strategic conversations with people who want repository orientation treated as durable infrastructure, not a one-off feature.
