The case for local

When you use a coding agent, it reads your files, runs your commands, and sees your prompts. That's exactly the data a memory layer needs to capture. The question is where that data goes after it's captured.

Cloud memory services send it to a server you don't control. Termyte keeps it in a SQLite file you do control.

# Your database is one file
./termyte.db

You can copy it, back it up, sync it, or ignore it. It's yours.

No API key required

A lot of memory tools assume you'll sign up for a hosted embedding API. Termyte runs embeddings locally through ONNX (Transformers.js). Nomic Embed v1.5 is the default; BGE Small is the fallback. No API key, no rate limits, no per-query cost.

termyte search "how does authentication work"

That query never leaves your machine.

The one optional call

Memory synthesis — turning raw traces into clean memories — benefits from an LLM. Termyte doesn't ship one. Instead, it reuses the LLM plan you're already paying for through your coding agent (Claude Code, Codex, OpenCode, or Gemini CLI). The call goes to the same provider you already use, not a separate Termyte service.

You can even turn synthesis off and still capture traces and run keyword search. The memory layer is useful at every level.