research · the lab

What we're building next.

We don't pre-announce products, but we can talk about the open problems that get us out of bed. Six research areas, ranging from "shipped today" to "long horizon."

Shipping · Genesis

Persistent AI memory

Local-first code-graph memory layer for AI coding agents. Symbol resolvers for Rust, TypeScript, Python. Symbol-anchored embeddings. The keystone product, shipping today as Mneme v0.4.0 Genesis.

SQLite ONNX tree-sitter BGE-small
Prototype

Structural certainty at scale

Real symbol resolvers across more languages. Cross-repo call graphs that survive renames and refactors. The goal: 10/10 canonical recall on the golden benchmark, then double the language coverage.

rustc-driver Pylsp ts-server graph-db
Research

Local-first LLM routing

When should an AI host route to a local 3B/7B model vs a cloud frontier model? Latency, cost, privacy, sensitivity of the question — all inputs. We're building the runtime decision engine that makes this automatic.

llama.cpp Phi-3 routing latency
Research

Cross-project federation

Code-graph signatures hashed and shared between machines you control. Lets your AI answer "have we seen this pattern before?" across every project you own — without ever sending source code anywhere.

blake3 privacy-first P2P
Research

Multimodal recall

Architecture diagrams, screenshots, voice notes — all indexed and recallable alongside code. Whisper for voice, image embeddings for visuals, all stitched into the same memory graph.

Whisper vision multimodal
Research

AI-native IDE primitives

What does an editor look like when AI is a first-class collaborator with persistent memory and structural understanding? Not a sidebar — a co-author with its own workspace.

MCP agent-native UX
open questions

Six open questions, if you want to talk.

If any of these are on your mind, we want to hear from you.

  1. What does persistent memory look like for an AI agent that lives a hundred sessions long?
  2. Can structural certainty fully replace text matching, or do we always need the fallback?
  3. How do you make federation private enough that engineers actually turn it on?
  4. What's the right boundary between local-first compute and cloud-first models?
  5. Is the right unit of AI memory a file, a symbol, or a decision?
  6. When the AI co-author has its own memory, who owns the institutional knowledge?

→ [email protected]