Memora cuts agent memory token use by 98%
Microsoft's open-source Memora compresses agent memory into brief abstractions, cutting context token use by 98% and setting new highs on two benchmarks.
Microsoft Research has open-sourced Memora, a memory framework for AI agents that need to operate across sessions measured in hours or days rather than turns. Published July 3 at ICML 2026 and available on GitHub under an MIT license, it targets the core problem with long-running agents: feeding full session history into the context window is expensive and eventually impossible, but dropping that history means the agent forgets.
The approach: Memora compresses each stored episode into a 6-to-8-word abstract representation. Retrieval uses those abstractions as keys; the rich underlying content surfaces only when a key matches the query. In tests, this cuts context token consumption by up to 98% compared to full-context methods.
On two standard benchmarks, LoCoMo (long-conversation memory) and LongMemEval (long-session question answering), Memora reaches new state-of-the-art results against Mem0, RAG-based memory, and full-context baselines.
What it means for you
The token cost of giving an agent real memory across sessions has been a practical ceiling for most teams. A 98% reduction, if it holds on your workload, changes the economics of multi-session agents: persistent customer history, ongoing project context, decisions the agent needs to recall from a week ago all become cheaper to support.
The code is public and the paper is through ICML 2026, so your team can evaluate it without waiting for a packaged product. Worth noting: research on memory tools can cut both ways, with some retrieval approaches hurting accuracy more than they help. Running Memora on your own tasks before committing is still the right first step. If you are building an agent that needs to operate across long sessions, browser-capable agents is a useful companion read on the operational side of that problem.