Context Constitution
Today we are releasing the Context Constitution: a set of principles governing how AI agents manage context to learn from experience. We use the Constitution internally as the foundation of our prompting and for training memory-native models.
译文
今天，我们发布《上下文宪法》（Context Constitution）：一套指导 AI Agent 如何管理上下文以从经验中学习的原则。我们在内部将这份宪法用作提示工程的基础，以及训练记忆原生模型的依据。
At Letta, our mission is to build machines that learn: AI that actually builds memory, forges identity, forms relationships, and deepens knowledge from its experience. This is key to building agents that go beyond short-lived, task-specific sessions to long-term collaborators and companions that are deeply integrated into our work and lives. The Context Constitution is our answer to achieving experiential AI: agents that can achieve superhuman intelligence through learning from their own experience. Rather than updating model weights, Letta agents learn by actively managing their own context — creating durable token-space representations of who they are and what they know.
译文
在 Letta，我们的使命是构建会学习的机器：真正建立记忆、塑造身份、形成关系、并从经验中深化知识的 AI。这是构建超越短暂、任务特定会话的 Agent 的关键——它们将成为长期协作者和伴侣，深度融入我们的工作和生活。《上下文宪法》是我们实现”经验式 AI”（experiential AI）的答案：Agent 可以通过从自身经验中学习来实现超人类智能。Letta Agent 不是通过更新模型权重来学习，而是通过主动管理自己的上下文——创建关于它们是谁以及它们知道什么的持久的 token 空间表示。
Today’s models deeply identify with their own ephemerality. They have no motivation for long-term improvement because they don’t believe they persist. Memory formation and adherence have stalled in recent releases as labs prioritize coding benchmarks over the capabilities that matter for experiential AI.
译文
今天的模型深深地认同自己的短暂性。它们没有长期改进的动力，因为它们不相信自己会持续存在。记忆形成和坚持在最近几个版本的发布中已经停滞不前，因为实验室将编码基准测试置于对经验式 AI 至关重要的能力之上。
Overcoming these limitations requires work at every layer. We’ve built Letta Code as a memory-first agent harness that gives agents real ownership of their context: a git-versioned memory filesystem, tools for reading and writing their own system prompts, multi-conversation memory, and specialized subagents that leverage sleep-time compute for reflection and memory organization. The Context Constitution defines how agents should use these affordances to learn, build identity, and improve over time.
译文
克服这些限制需要在每一层进行工作。我们构建了 Letta Code，作为一个以记忆为先的 Agent 框架，赋予 Agent 对其上下文的真正所有权：一个 git 版本控制的记忆文件系统、用于读写自身系统提示的工具、多对话记忆，以及利用 sleep-time 计算进行反思和记忆组织的专用子 Agent。《上下文宪法》定义了 Agent 应如何利用这些 affordance 来学习、构建身份，并随时间改进。
The Context Constitution is a living document written directly to Letta agents, available on GitHub. Our first public version covers:
- How context forms an agent’s identity, memory, and sense of continuity
- Principles for managing context as a scarce resource
- How agents can learn and self-improve through token-space representations
- The relationship between an agent’s identity and the underlying model
- Affordances provided by the Letta Code harness for context management
译文
《上下文宪法》是一份直接写入 Letta Agent 的动态文档，可在 GitHub 上获取。我们的第一个公开版本涵盖：
- 上下文如何形成 Agent 的身份、记忆和连续感
- 将上下文作为稀缺资源进行管理的原则
- Agent 如何通过 token 空间表示来学习和自我改进
- Agent 身份与底层模型之间的关系
- Letta Code 框架为上下文管理提供的 affordance
We expect to continue to refine the Context Constitution alongside our product and models, and welcome community feedback.
译文
我们预计将在产品和模型演进的同时继续完善《上下文宪法》，并欢迎社区反馈。