Mymir

Context management for the agent-native engineering era.

Most of us aren't really writing code anymore, we're directing agents that do. But those agents have no memory. Every session starts from zero, and engineers end up spending their time re-explaining what was built, why decisions were made, and what still needs to happen. That's not engineering, that's babysitting.

Mymir replaces that cycle. It's not just a context layer your agents read from, it's an end-to-end project management tool that agents operate natively. Mymir creates tasks, refines them, plans implementations, provides the right context at the right stage, and tracks everything that happens. Your agent harness doesn't need a briefing. It walks into every session knowing exactly what to do next and why.

How it works

Instead of docs, wikis, or messy markdown files, Mymir treats project context as a live knowledge base agents can reason from.

We built Mymir around two core concepts:

Context network. A living map of your project that captures not just what was built, but why decisions were made, what was tried and abandoned, and how different parts of the codebase relate to each other.

Context retrieval interface. Four context shapes, one per job. Each is arranged by U-shaped attention (highest-recall content at the start and end) so what matters most lands where LLMs read best:

Together, they don't just inform your agent, they drive it. Mymir manages the full lifecycle: Brainstorm > Decompose > Refine > Plan > Execute > Track.

Describe your idea and Mymir decomposes it into tasks with dependency edges, determines what's ready to plan or implement, and hands your agent the exact context it needs for that stage. When a task is plannable, your agent gets the spec, prerequisites, and related work. When it's ready to implement, your agent gets the full execution context: upstream decisions, file paths, and acceptance criteria.

The agent moves from task to task with the right context at every step, no manual handoff required.

We're building Mymir using Mymir, so everything described here is something we're living in real time.

How it looks

The web UI has two modes: Structure and Graph.

Structure mode puts your task list on the left and a detail panel on the right. You refine specs, track progress, and review execution records without switching views.

Graph mode overlays the context network so you can see how tasks, decisions, and dependencies connect while still working in the detail panel.

Zoom out and the full graph renders your entire context network. Clusters, bottlenecks, and orphaned work become obvious at a glance.

How it runs

Mymir ships as a Next.js web app plus vendor-native plugins for Claude Code, Codex

... View full README on GitHub