← 回總覽

IBM 推出“轨迹感知记忆”技术,解决 AI 智能体遗忘问题

📅 2026-03-14 18:30 God of Prompt 人工智能 3 分鐘 3676 字 評分: 84
AI 智能体 IBM 研究院 轨迹感知记忆 提示工程 机器学习
📌 一句话摘要 IBM 新推出的“轨迹感知记忆”技术,能让 AI 智能体无需重新训练即可从过往执行中学习,显著提升其在复杂任务上的表现。 📝 详细摘要 IBM 开发了一种名为“轨迹感知记忆”(Trajectory-Informed Memory)的解决方案,旨在解决 AI 智能体的“遗忘”问题。通过在执行过程中捕获哪些方法有效、哪些失败以及哪些效率低下,这些洞察作为可复用经验被注入到未来的提示词中。该方法使复杂的、多步骤工作流的完成率(从 19.1% 提升至 47.6%)相对提高了 149%,且无需进行任何模型重新训练。 📊 文章信息 AI 评分:84 来源:God of Prompt(
Skip to main content ![Image 1: LogoBestBlogs](https://www.bestblogs.dev/ "BestBlogs.dev")Toggle navigation menu Toggle navigation menuArticlesPodcastsVideosTweetsSourcesNewsletters

⌘K

Change language Switch ThemeSign In

Narrow Mode

IBM Introduces Trajectory-Informed Memory for AI Agents =======================================================

IBM Introduces Trajectory-Informed Memory for AI Agents ======================================================= ![Image 2: God of Prompt](https://www.bestblogs.dev/en/tweets?sourceId=SOURCE_f58696ab) ### God of Prompt

@godofprompt

🚨 BREAKING: IBM just admitted your AI agent forgets everything the moment it finishes a task.

> Every mistake. Repeated.

> Every inefficiency. Repeated.

> Every failure. Repeated.

They built the fix.

> Every AI agent starts each task from zero:

> No memory of what worked.

> No memory of what failed.

> No memory of the faster path it found yesterday.

IBM built a fix called Trajectory-Informed Memory.

It watches the agent's full execution and extracts three types of reusable tips:

> what worked

> what failed and how it recovered

> what succeeded but wasted steps

Those tips get injected into the agent's prompt next time a similar task appears.

The model stays frozen. No retraining.

Only the memory evolves.

> 14.3 pp gain in scenario completion on tasks never seen before

> Complex tasks: 19.1% → 47.6% scenario completion, a 149% relative increase

> Zero retraining required

The 149% on hard tasks is the number.

These are 50+ step workflows across multiple apps. Exactly where agents break in production.Show More

!Image 3: Tweet image

Mar 14, 2026, 10:30 AM View on X

1 Replies

0 Retweets

3 Likes

475 Views G God of Prompt @godofprompt

One Sentence Summary

IBM's new Trajectory-Informed Memory allows AI agents to learn from past executions without retraining, significantly improving performance on complex tasks.

Summary

IBM has developed a solution called Trajectory-Informed Memory to address the 'forgetfulness' of AI agents. By capturing what worked, what failed, and what was inefficient during execution, these insights are injected into future prompts as reusable tips. This method resulted in a 149% relative increase in completion rates for complex, multi-step workflows (from 19.1% to 47.6%) without requiring any model retraining.

AI Score

84

Influence Score 2

Published At Today

Language

English

Tags

AI Agents

IBM Research

Trajectory-Informed Memory

Prompt Engineering

Machine Learning HomeArticlesPodcastsVideosTweets

IBM Introduces Trajectory-Informed Memory for AI Agents |... ===============

查看原文 → 發佈: 2026-03-14 18:30:52 收錄: 2026-03-14 20:00:53

🤖 問 AI

針對這篇文章提問,AI 會根據文章內容回答。按 Ctrl+Enter 送出。