Splitting an AI agent’s memory into routed clusters reduced estimated API cost by 65–72% against a single shared history, in tests across four language models, according to a new paper from six researchers at FAIR at Meta.
The method, Routed Continual Learning, also raised normalized reward over the single global history by 15–30% of the maximum attainable reward. When tasks were interleaved, the global history gained almost nothing over no memory at all: a mean of +0.015, against 0.280 for the routed method, about 18.7 times as much, by our arithmetic.
Short explainer · Meta’s Routed Agent Memory Cut API Cost Up to 72%
A vertical news Short summarising this story. Figures come from the FAIR at Meta paper on arXiv; our arithmetic is labelled.
Business Pill · The Memory Wall
An evergreen Business Engineer explainer on memory as a constraint in AI systems. It teaches the general idea only and does not discuss this paper.
The key insight: In the FAIR at Meta paper, routing an agent’s past runs into separate memory clusters cut estimated API cost by 65–72% against one shared history, so the routed method cost about a third, by our arithmetic.
The paper groups past agent runs into clusters, and each cluster keeps its state in one memory system chosen from a fixed library of five. Before each run, a router, “implemented as a large language model”, decides whether to reuse an existing cluster or create a new one, the paper says.
The router sees only the current task, the memory system descriptions and the cluster summaries, and task labels are not supplied during evaluation. A second component, Context Memory Adapter, “mediates between the memory system and the agentic loop”.
Mean gain in normalized reward over no memory for three memory set-ups, and estimated API cost of the routed method against one shared history, on a 398-task interleaved stream across four models. Figures from the FAIR at Meta paper on arXiv; cost range by our arithmetic; chart by FourWeekMBA.
What Was Posted on X
Brilliant paper on agent memory from FAIR at Meta.
(bookmark it)
A single global history is the usual way to let agents learn from past runs. When tasks are interleaved, it gains almost nothing over no memory at all, an average of +0.015 in normalized reward.
This work proposes Routed Continual Learning. Past trajectories are grouped into clusters, and each cluster keeps its state in one memory system from a fixed library. Before each run, an LLM router reuses a cluster or creates a new one.
They tested it on a stream mixing 301 CL-Bench instances with 97 banking instances from ฯยณ-Bench, across four language models.
Routed CL raised reward over the global history by 15% to 30% of the maximum attainable reward. Estimated API cost fell by 65% to 72%.
The runs were sequential, so parallel agents updating one cluster remain untested.
If your agents handle mixed work, it might help to split their memory into clusters before adding a bigger memory system.
DAIR.AI (@dair_ai) flagged the paper on X on 31 October and summarised its main results, including the 65% to 72% fall in estimated API cost. The figures in this piece come from the paper itself.
What the Tests Showed
The team tested a stream of 398 tasks that interleaves 301 instances from the six CL-Bench tasks with 97 banking instances from τ³-Bench, the paper says. The four models were Claude Opus 4.7, Claude Opus 5, GPT-5.6 Sol and Gemini 3.7 Flash, each run five times.
The paper says Routed CL achieved higher reward than both the single global history and router-partitioned histories for all four models. It notes that the original CL-Bench had found the plain shared history strongest when tasks were evaluated as separate streams. A 65–72% cut means the routed method cost about a third of the global history, by our arithmetic.
The Business Lever: Memory as a Cost Line
For companies running agents on mixed work, the paper frames memory design as a routing problem that decides “how experience is represented and which instances share state”. The cost result is an estimate of API spend, not a measured bill.
As we read the paper, the result is about mixed work: one shared history did well on separate task streams in the original benchmark, but gained almost nothing once tasks were interleaved.
As we read it, the cost and the reward moved in the same direction, since the paper reports higher reward for all four models at lower estimated cost.
As we read it, the open question is scale: the runs were sequential, so agents updating the same memory cluster at once have not been tested.
Philip et al., arXiv abstract, 28 Oct 2026
“These results support approaching continual learning as a routing problem that chooses both a learning mechanism and a state cluster rather than committing to one representation globally.”
Three Implications
FOR AGENT BUILDERS The paper tested a router that picks a memory system and a cluster for each run, using only the task, memory descriptions and cluster summaries.
FOR FINANCE TEAMS The paper reports estimated API cost 65–72% lower than a single global history across four models.
FOR MODEL PROVIDERS The paper ran the same comparison on Claude Opus 4.7, Claude Opus 5, GPT-5.6 Sol and Gemini 3.7 Flash, and reports the routed method ahead for all four.
The framework says “The new job of finance is to manage the ratio between the two — work produced per token burned.”
As we read it, the paper reports more reward for less estimated spend, which is that ratio moving on both sides. The framework is the owner’s argument and does not discuss this paper or its figures.
What Is Not Established
The paper says its experiments “process instances sequentially”, and that it does “not evaluate simultaneous updates, stale reads, or cross-model reuse of cluster state”. It calls these important tests for distributed deployment. The results come from benchmark streams, and the paper is a preprint posted on arXiv.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
FAIR at Meta researchers report that routing an agent’s past runs into separate memory clusters raised reward and cut estimated API cost by 65–72% against one shared history, across models from Anthropic, OpenAI and Google. Agents updating the same cluster in parallel remain untested.
A note on sourcing. All facts and figures come from the paper by Philip and colleagues at FAIR at Meta on arXiv (28 October 2026) and its abstract, read on 31 October; the results are the authors’. We did not contact the authors or Meta.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
