Research

The open problems.

We do not announce products early. We are glad to discuss the open problems that occupy us. There are six, and they run from shipping today to years away.

See the six areas Read the open questions

Move the pointer to look into the dark. What you look at wakes.

The lab

Six research areas.

01of 06

Persistent AI memory

A local-first code-graph memory for AI coding agents, with symbol resolvers for Rust, TypeScript, and Python and embeddings anchored to symbols. It ships today as Mneme v0.4.1.

Status
Shipping · Genesis
Symbol resolvers
  • Rust
  • TypeScript
  • Python
Keywords
  • SQLite
  • ONNX
  • tree-sitter
  • BGE-small

02of 06

Structural certainty at scale

Symbol resolvers for more languages, and call graphs across repositories that survive a rename. The goal is 10 out of 10 canonical recall on our golden benchmark.

Status
Prototype
Goal
10 / 10 canonical recall. A goal, not yet a result.
Keywords
  • Rust resolver
  • TypeScript resolver
  • Python resolver

03of 06

Local-first LLM routing

Deciding when an AI host should send a task to a local model of 3 or 7 billion parameters, and when to a frontier model in the cloud. Latency, cost, and privacy all bear on the answer. We are building the engine that decides at run time.

Status
Research
Keywords
  • llama.cpp
  • Phi-3
  • routing

04of 06

Cross-project federation

Signatures of a code graph, hashed and shared only between machines you control. Your AI can then say whether a pattern has been seen before, and no source code leaves the machine.

Status
Research
Keywords
  • hashed signatures
  • privacy-first
  • P2P

05of 06

Multimodal recall

Architecture diagrams, screenshots, and voice notes, indexed and recalled beside the code, in the same memory graph.

Status
Research
Keywords
  • Whisper
  • vision
  • multimodal

06of 06

AI-native IDE primitives

What an editor becomes when AI is a full collaborator with persistent memory and structural understanding: a co-author with a workspace of its own, where today it has a sidebar.

Status
Research
Keywords
  • MCP
  • agent-native
  • UX

Open questions

Six questions we cannot yet answer.

If one of these is yours too, write to us.

[email protected]
  1. What should persistent memory be for an AI agent that lives for a hundred sessions?
  2. Can structural certainty replace text matching entirely, or will a fallback always be needed?
  3. How private must federation be before engineers will switch it on?
  4. Where does the boundary fall between local-first compute and cloud-first models?
  5. What is the proper unit of AI memory: a file, a symbol, or a decision?
  6. When an AI co-author has a memory of its own, who owns the institutional knowledge?