Trajectory: Agent 经验数据的标准格式

Letta 2026-07-25 5.5 KB 209.4K 748 𝕏 文章列表

文章介绍了 Trajectory，一个开源的数据格式标准化工具，旨在统一不同编码代理会话（如 Claude Code、Codex 等）的数据格式。该格式注重令牌效率，便于代理从过去的经验中学习，解决跨平台数据碎片化问题。

# Trajectory: A Standard Format for Agent Experience Data **作者**: Letta **日期**: 2026-07-24T06:07:03.000Z **来源**: [https://x.com/Letta_AI/status/2080535211473850822](https://x.com/Letta_AI/status/2080535211473850822) --- ![Image](/doc/3420/image/img_640038cab6f3.jpg) Introducing trajectory, an open-source package that normalizes coding-agent sessions from @AnthropicAI Claude Code, @OpenAI Codex, @pidotdev, @LangChain deepagents, @openclaw, Letta Code, and other harnesses into one token-efficient format designed for agents learning from past experience. Agents today have a limited ability to learn from past experience to improve in the future. Furthermore, agent experience is split across many different harnesses: many users swap between harnesses such as Claude Code, Codex, and Letta Code - and organizations typically have individual users on a wide variety of harnesses depending on personal preference. Creating memories, learning, or dreaming therefore is limited to a single harness, and a single user. Learning across harnesses requires first creating a standard data format for the experience each harness produces. Today, learning is primarily done in token-space, where agents themselves process experience to learn system prompts, skills, or even harness-level modifications. An ideal data format for experience is therefore: 1. Standardized across harnesses, to avoid distraction from divergence in formats 2. Token-efficient, to enable agents to process experience as efficiently as possible While projects such as Harbor have introduced normalized formats for agent trajectories (ATIF), those formats are intended for full-fidelity replay and benchmarking - they preserve per-step token metrics, structured tool payloads, and untruncated outputs. The trajectory package formats trajectories across harnesses in a consistent, agent-friendly format. The format is designed for agents reading past sessions for memory formation, and minimizes token overhead from metadata. ## Trajectory schema A trajectory is a list of records in a standard format: assistant messages, user messages, reasoning, tool calls, and tool results. The first record contains metadata about the harness the trajectory was created in. ![Image](/doc/3420/image/img_86fee3e8af0b.jpg) The full spec is defined in trajectory-v1.schema.json. The format keeps only the information needed to understand the agent's experience. It drops harness bookkeeping (per-line envelopes, duplicated payloads, UI event streams, encrypted reasoning blobs), and optionally truncates long tool results. On sessions we sampled, this results in a ~5x reduction in token counts compared to native session formats. ## Listing & formatting trajectories The trajectory package lets you list sessions from any harness and convert them into normalized trajectories. You can use the package to aggregate data across harnesses for indexing, or for processing with memory agents. ``` import { listTrajectories, normalizeTranscript } from "@letta-ai/trajectory"; import { readFileSync } from "fs"; // Discover local Claude Code sessions (also works for codex, letta-code, ...) const page = await listTrajectories({ source: "claude-code", limit: 10 }); // Normalize one into the standard record format const transcript = readFileSync(page.items[0].path, "utf8"); const { records, diagnostics } = normalizeTranscript({ source: "claude-code", transcript, }); ``` # How we're using trajectory The Letta Code harness now uses the trajectory format to normalize session data from other local agents like Claude Code and Codex. Past trajectory data is reviewed to bootstrap the agent's memory. With the trajectory format, agent dreaming can select and normalize sessions from every harness on the machine, so lessons learned in a Claude Code or Codex session inform the agent's memory the same way its own sessions do. We're exited about how aggregating experience across many harnesses (or even many users) can translate into learning and memory that makes better agents. ## Next steps Try out the package: ``` npm install -g @letta-ai/trajectory ``` You can also run `/init` in Letta Code to review experience from other harnesses: ``` npm install -g @letta-ai/letta-code ``` Read this post in its original format: https://www.letta.com/blog/trajectory/ ## 相关链接 - [Letta](https://x.com/Letta_AI) - [@Letta_AI](https://x.com/Letta_AI) - [84K](https://x.com/Letta_AI/status/2080535211473850822/analytics) - [trajectory](https://github.com/letta-ai/trajectory) - [@AnthropicAI](https://x.com/@AnthropicAI) - [@OpenAI](https://x.com/@OpenAI) - [@pidotdev](https://x.com/@pidotdev) - [@LangChain](https://x.com/@LangChain) - [@openclaw](https://x.com/@openclaw) - [dreaming](https://www.letta.com/blog/sleep-time-compute/) - [token-space](https://www.letta.com/blog/continual-learning/) - [agent trajectories (ATIF)](https://github.com/harbor-framework/harbor/blob/main/rfcs/0001-trajectory-format.md) - [trajectory](https://github.com/letta-ai/trajectory) - [trajectory-v1.schema.json](https://github.com/letta-ai/trajectory/blob/main/schema/trajectory-v1.schema.json) - [https://www.letta.com/blog/trajectory/](https://www.letta.com/blog/trajectory/) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [2:07 PM · Jul 24, 2026](https://x.com/Letta_AI/status/2080535211473850822) - [84.5K Views](https://x.com/Letta_AI/status/2080535211473850822/analytics) - [View quotes](https://x.com/Letta_AI/status/2080535211473850822/quotes) --- *导出时间: 2026/7/25 11:03:17* --- ## 中文翻译 # Trajectory：一种代理经验数据的标准格式 **作者**: Letta **日期**: 2026-07-24T06:07:03.000Z **来源**: [https://x.com/Letta_AI/status/2080535211473850822](https://x.com/Letta_AI/status/2080535211473850822) --- ![Image](/doc/3420/image/img_640038cab6f3.jpg) 隆重介绍 trajectory，这是一个开源包，旨在将来自 @AnthropicAI Claude Code、@OpenAI Codex、@pidotdev、@LangChain deepagents、@openclaw、Letta Code 以及其他工具的编码代理会话规范化为一种对 Token 友好的统一格式，专为代理从过往经验中学习而设计。如今，代理从过往经验中学习以在未来改进的能力十分有限。此外，代理经验分散在许多不同的工具中：许多用户在 Claude Code、Codex 和 Letta Code 等工具之间切换使用——而组织内的不同用户通常根据个人偏好使用各种各样的工具。因此，创建记忆、学习或“做梦”往往局限于单一工具和单一用户。要实现跨工具的学习，首先需要为每个工具产生的经验创建一种标准数据格式。目前，学习主要在 Token 空间中进行，代理在其中自行处理经验，以学习系统提示词、技能，甚至是工具层面的修改。因此，理想的经验数据格式应当： 1. 跨工具标准化，以避免因格式差异而分散注意力 2. Token 高效，以使代理能够尽可能高效地处理经验尽管 Harbor 等项目已经引入了代理轨迹的规范化格式（ATIF），但这些格式旨在用于全保真回放和基准测试——它们保留了每步的 Token 指标、结构化工具负载和未截断的输出。trajectory 包则将跨工具的轨迹格式化为一致的、对代理友好的格式。该格式专为读取过往会话以形成记忆的代理而设计，并将元数据的 Token 开销降至最低。 ## Trajectory 架构 Trajectory 是标准格式的记录列表：包含助手消息、用户消息、推理、工具调用和工具结果。第一条记录包含创建该轨迹的工具的元数据。 ![Image](/doc/3420/image/img_86fee3e8af0b.jpg) 完整的规范定义在 trajectory-v1.schema.json 中。该格式仅保留理解代理经验所需的信息。它舍弃了工具的簿记信息（每行信封、重复负载、UI 事件流、加密推理块），并可选择截断过长的工具结果。在我们抽样的会话中，与原生会话格式相比，这使 Token 数量减少了约 5 倍。 ## 列出与格式化轨迹 trajectory 包允许您列出任何工具的会话并将其转换为规范化轨迹。您可以使用该程序包聚合跨工具的数据以进行索引，或供记忆代理处理。 ``` import { listTrajectories, normalizeTranscript } from "@letta-ai/trajectory"; import { readFileSync } from "fs"; // 发现本地 Claude Code 会话（也适用于 codex、letta-code 等） const page = await listTrajectories({ source: "claude-code", limit: 10 }); // 将其中一个规范化为标准记录格式 const transcript = readFileSync(page.items[0].path, "utf8"); const { records, diagnostics } = normalizeTranscript({ source: "claude-code", transcript, }); ``` # 我们如何使用 trajectory Letta Code 工具现在使用 trajectory 格式来规范化来自其他本地代理（如 Claude Code 和 Codex）的会话数据。过去的轨迹数据会被审查以引导代理的记忆。有了 trajectory 格式，代理的“做梦”过程可以选择并规范化机器上每个工具的会话，因此在 Claude Code 或 Codex 会话中学到的教训，会像代理自己的会话一样为代理的记忆提供信息。我们对将许多工具（甚至许多用户）的经验聚合起来，转化为能够打造更优秀代理的学习和记忆感到兴奋。 ## 后续步骤试用该包： ``` npm install -g @letta-ai/trajectory ``` 您也可以在 Letta Code 中运行 `/init` 来审查来自其他工具的经验： ``` npm install -g @letta-ai/letta-code ``` 阅读原文格式的文章： https://www.letta.com/blog/trajectory/ ## 相关链接 - [Letta](https://x.com/Letta_AI) - [@Letta_AI](https://x.com/@Letta_AI) - [84K](https://x.com/Letta_AI/status/2080535211473850822/analytics) - [trajectory](https://github.com/letta-ai/trajectory) - [@AnthropicAI](https://x.com/@AnthropicAI) - [@OpenAI](https://x.com/@OpenAI) - [@pidotdev](https://x.com/@pidotdev) - [@LangChain](https://x.com/@LangChain) - [@openclaw](https://x.com/@openclaw) - [dreaming](https://www.letta.com/blog/sleep-time-compute/) - [token-space](https://www.letta.com/blog/continual-learning/) - [agent trajectories (ATIF)](https://github.com/harbor-framework/harbor/blob/main/rfcs/0001-trajectory-format.md) - [trajectory](https://github.com/letta-ai/trajectory) - [trajectory-v1.schema.json](https://github.com/letta-ai/trajectory/blob/main/schema/trajectory-v1.schema.json) - [https://www.letta.com/blog/trajectory/](https://www.letta.com/blog/trajectory/) - [升级到 Premium](https://x.com/i/premium_sign_up) - [2:07 PM · Jul 24, 2026](https://x.com/Letta_AI/status/2080535211473850822) - [84.5K 次查看](https://x.com/Letta_AI/status/2080535211473850822/analytics) - [查看引用](https://x.com/Letta_AI/status/2080535211473850822/quotes) --- *导出时间: 2026/7/25 11:03:17*