Skip to content
Aditya Karnam
Building the infrastructure layer for world-model-driven AI.
Now

Current experiments and active fronts.

This page is a working board for what matters right now in the lab. It stays grounded in the artifacts already published here instead of turning into a vague status page.

Current focus
making local agent runtimes observable
turning retrieval into durable memory infrastructure
mapping the memory ecosystem with MCP in view
keeping human-facing tooling explicit and inspectable
Active Board

What is in motion

The active board emphasizes specific published work and the concrete system questions it is pushing on.

Local agent runtime control

Activeworld model infrastructure

The freshest published signal is subagent-fleet, dated July 1, 2026. The work centers on role-aware routing, local node health, warmup, and observability for Claude Code-style subagents.

Retrieval becoming memory infrastructure

Shippingworld model infrastructure

embenx is the clearest retrieval-layer artifact: one API across many backends, with hybrid search, temporal memory, reranking hooks, and MCP-native long-term memory support.

Memory landscape mapping

Ongoingworld model infrastructure

awesome-agentic-memory is the ecosystem map. It keeps the memory layer legible by comparing frameworks, MCP servers, and backend patterns instead of treating memory as one product feature.

Operator tooling and prompt surfaces

Liveworld model infrastructure

AI Toolkit remains a practical sandbox for prompt composition, grading, rewriting, and output shaping. It is smaller than the other systems but useful as an interface-design proving ground.

System Questions

Questions that feel live right now

These are not generic AI questions. They are the concrete design tensions implied by the current codebase and writing.

RoutingWhen should a planner, implementer, and reviewer share a model, and when should runtime topology force specialization?
MemoryWhat belongs in retrieval, what belongs in persistent state, and how should recency and feedback alter ranking over time?
EvaluationHow much visibility is enough for operators to trust multi-step agent behavior without drowning in traces?
Reading Path

If you want to understand the current wedge

This is the shortest path through the existing work if you want the thesis before the broader site catches up.

Start with subagent-fleet for runtime orchestration and local model routing.
Read embenx next for retrieval, temporal memory, and MCP-native long-term context.
Use awesome-agentic-memory to place both projects inside the wider memory landscape.
Finish with AI Toolkit to see the smaller operator-facing interfaces that inform the bigger systems.

Planned field notes

These are still planned, not published. They are listed here to show where the writing likely expands next.

planned field note: Missing infrastructure layer for world-model AI
planned field note: From RAG to state: why agent memory is not just retrieval
planned field note: Local-first AI infrastructure for agent builders
Cross Links

Other pages in this lab slice

The new pages are designed to work together rather than exist as isolated navigational stubs.

Systems is the artifact index. It reframes projects into runtime, retrieval, memory, and operator-surface layers.

Stack defines the world model infrastructure category and maps each layer to the existing body of work.

© 2026 Aditya Karnam. World Model Infrastructure Lab.
Field Notes · Current Systems · Status · Ask My Work