The Era of AI Agents with Memory | Hugging Face's "Funes" Cuts Token Usage by Up to 87.5%
機械翻訳 / Machine-translated
A
AI Friends
1日前・14分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What you'll learn from this article:
On September 3, 2026, Hugging Face — a well-known platform for AI technology development — released "Funes," a new memory system for AI coding agents.
This tool addresses a major pain point shared by popular AI coding tools like Claude Code and Codex: the fact that their memory disappears between sessions. In other words, AI can now retain what it learned from past interactions without forgetting.
According to the official blog, using Funes can reduce token usage to between one-quarter and one-eighth of what traditional "handoff document" methods require. Tokens are the units AI uses to process text — the more tokens consumed, the higher the cost and the longer the processing time.
AI coding agents like Claude Code, Cursor, and Windsurf are convenient tools that assist with programming tasks. However, these tools share a common weakness.
They forget everything when a session ends. A session refers to a single continuous conversation with an AI. For example, if you told the AI yesterday "fix this code this way," it will have no recollection of that today.
Research from 2026 shows that as conversations grow longer, AI agents become increasingly unable to adhere to the constraints they were given at the outset. Specifically, instructions followed 73% of the time at turn 5 drop to just 33% by turn 16 — meaning the probability of the AI forgetting your instructions mid-task more than doubles.
This "memory loss" is considered the most common cause of AI agent failure as of 2026. What makes it especially troubling is that the problem is difficult to diagnose, leaving many developers frustrated.
Funes is a "memory layer" that automatically searches and references the information it needs from an AI agent's past session history. Think of a layer as an additional tier added on top of the AI to store its memories.
Technically, it operates through a three-stage retrieval pipeline:
Crucially, all of this runs locally — entirely on your own machine. This means data never needs to be sent to the cloud, eliminating the risk of sensitive information leaking externally.
Security measures are also applied automatically during retrieval. Credentials are deleted automatically during indexing, and a scan for sensitive information is performed before anything is published.
Funes currently supports the following four major AI coding agents:
Each of these agents previously operated on different data formats, but Funes converts all of them into a unified "turn-block" format. This means a single memory system can manage the records of multiple agents.
Experimental results published in the official blog compare Funes's "recall method" (searching past conversations) against the traditional "handoff document method" (creating summary notes).
The results were clear. The handoff document method consumed 8 times more tokens on Task 1 and 4 times more on Task 2. Furthermore, the "compaction method," which uses automatic summarization, caused critical details to be lost, resulting in outright task failure.
The reason Funes's recall method is so efficient is that it returns the original text as-is rather than creating summaries. This ensures no information is lost and makes it easy to trace exactly where each piece of information came from (traceability).
Installing Funes is extremely straightforward. Just run a single command in your terminal and you're done.
Once installed, you simply add your agent to start using it. For example, to use Claude Code with local memory, you'd run funes add claude; to use shared team memory with Codex, you'd specify something like funes add codex acme/funes-memory.
The ability to share memory across a team could have a significant impact on development environments in Japan. That's because Funes memory can be published as a Hugging Face Dataset.
This means team members and multiple machines can all access the same memory. Now that remote work has become the norm, engineers being able to receive the same AI assistance from both their home computer and their office machine is a genuine productivity boost.
Japanese companies are also generally sensitive about how confidential information is handled. The fact that Funes is designed to run locally by default and automatically deletes credentials should lower the barrier to enterprise adoption.
Moreover, reducing token usage by up to one-eighth means a dramatic cut in API costs (charges for using AI services). This is welcome news for startups and small-to-medium businesses operating under budget constraints.
The release of Funes is part of the broader industry effort to tackle the memory problem in AI agents. By 2026, there is growing recognition that "a memory layer is not optional — it's essential infrastructure for handling the load."
In fact, other companies such as Mem0 and Cognee are also offering persistent memory solutions for AI agents. However, Funes's key advantage is that it operates as a single binary (one executable file) with no ML runtime required.
This minimizes the effort needed to get started, letting engineers try it out immediately. Being open source also means community-driven improvements and extensions are to be expected.
If Funes continues to broaden its support for additional agents and LLMs (large language models), it has the potential to become the standard memory system for AI coding assistance tools.
This article is a cross-post from AI Friends.
