Skip to main content

ICE: turn-granular memory for AI agents

ICE, the Indexless Context Engine, is turn-granular memory for AI agents that makes the context you already have useful.

It is not a vector database or a RAG pipeline. It is an embedded memory engine that you mount directly into your application, with a zero-config local mode.

How it works​

  • Turn-granular storage. Each conversational turn is stored as its own frame, so the order of events survives. Fixed-window chunking loses it.
  • Multi-lane retrieval. Retrieval is organised like the sections of a library: episodic for what happened, semantic for what it means, procedural for how to do it, and working for what the agent is focused on now.
  • Cross-session persistence. Memory carries across sessions, so each interaction can build on the last.

The useful window​

While others chase longer contexts, ICE asks what makes the information in the window useful. It works on the signal-to-noise ratio of the context an agent already has.

Core principles​

  1. Private & embedded: ICE runs inside your application process. There is no external server to run, and your data stays in your environment.
  2. Compiled core: The core memory operations run in a pre-compiled binary, bypassing the overhead of standard Python for critical tasks.
  3. Typed SDK: The SDK is type-safe, so the API is checked before you run it.

Next steps​