A reusable memory model experiment for LLM agents
The idea is to train one fixed memory model once and use it with different LLMs. This would only work well when the memory model and the chosen LLM share the same meaning. A separate small model would translate the between the two systems.
That translator is meant to be only tens of megabytes, or available from a shared . The idea combines inspiration from the Titans memory paper with RescursiveMAS work on training small models to translate s. The goal is better LLM memory without relying on , , or elaborate prompts.
The is open for others to test, break, disprove, or find errors in the current results.
Key points
- The project proposes one fixed memory model that can work with different LLMs.
- A small translator model would convert information between the memory model and the chosen LLM.
- The translator is intended to be only tens of megabytes.
- The approach avoids , , and complex as the main memory method.
- The is open for testing and criticism of the current results.