# AMA

AMA is a research system for AI agent memory that uses four collaborating agents and three memory granularities to store, verify and refresh what a conversational agent remembers over time.

AMA (Adaptive Memory via Multi-Agent Collaboration) addresses long-term memory for LLM agents by combining raw-text memory, atomic fact memory and event-level episode memory, and choosing which granularity best answers a given question. A Constructor agent builds the three memory forms from conversation, a Retriever rewrites the query and routes it to the right granularity, a Judge checks relevance and consistency and can request another retrieval round, and a Refresher updates or removes memories that conflict with a user's latest state.

The repository is the official research code for a paper accepted at the Findings of ACL 2026, and also packages a reusable memory skill with a reference adapter for the OpenClaw agent gateway.

It is written in Python and released under the Apache-2.0 license.

Website: https://sherlockwz.github.io/AMA/