Letta Code: Stateful Coding Agents That Learn and Lead on Terminal-Bench

Letta Code turns coding agents into persistent teammates that learn from your codebase, actions, and feedback. It supports memory initialization, explicit reflection, and reusable skills stored as simple .md files, with powerful search over past work. It also ranks #1 among model-agnostic OSS harnesses on Terminal-Bench, rivaling provider-specific tools.

Key Points

• Letta Code creates long-lived, stateful coding agents that learn from past interactions and project context.
• /init bootstraps memory by analyzing the local codebase and rewriting the agent’s system prompt via memory blocks; /remember triggers explicit reflection.
• Repeated tasks can be captured as reusable, versioned skills (.md files) that agents can share and reuse.
• Agents can search persisted conversations and tools via Letta’s API with vector, full-text, and hybrid search.
• Letta Code is the #1 model-agnostic OSS harness on Terminal-Bench, matching provider-specific harness performance and surpassing prior OSS baselines.

Sentiment

The overall sentiment is cautiously optimistic, marked by significant interest in Letta Code's technical approach to memory. While there's broad skepticism and negative experience with existing opaque memory systems (like ChatGPT's), Letta Code's transparent and controllable 'white box' memory system is perceived as a promising solution that addresses many of these concerns, leading to a generally positive and inquisitive reception.

In Agreement

• The value of agents learning and retaining context over time to avoid repetitive mistakes (e.g., specific `git add` patterns) is acknowledged.
• Letta's 'white box' memory system, which provides transparency and control over how memory influences prompts, is crucial for preventing 'context poisoning' experienced with opaque memory solutions like ChatGPT's.
• The ability to distill repeated workflows into reusable, versioned skill files (learned skills) is seen as an effective way to improve agent performance over time.
• Structured, project-specific context (like an `llm.md` file) is considered essential for LLMs to understand project quirks, avoid undesirable behavior, and manage unique coding requirements.
• Letta Code's strong performance on Terminal-Bench as the #1 model-agnostic open-source harness validates its effectiveness as a coding agent.

Opposed

• Skepticism about the general utility of 'memory' in coding agents, with some arguing that well-maintained documentation and feature specifications in a repository suffice.
• Concerns that maintaining memory could become a burden, leading to the accumulation of irrelevant or problematic information (e.g., trivial linting fixes) that degrades agent performance.
• The significant problem of 'context poisoning' where accumulated, often incorrect or outdated, memory in black-box systems can negatively impact an agent's responses and reliability.
• Philosophical worries that entrenched AI memory could make agents 'stuck in their ways' or resistant to adopting new behaviors, akin to human stubbornness, questioning if memory is a bug or a feature.