Searching over Session Transcripts > MEMORY.md (github.com)

A new tool called "agent-traces" has been released, enabling local semantic search across JSONL session transcripts generated by AI coding assistants like Claude Code and Codex. Users can now easily recall past prompts and tool outputs from thousands of conversation logs stored on their machines, enhancing the capability of AI agents to draw from previous interactions. By building a semantic index, "agent-traces" allows agents to reference prior work, making it possible to improve the efficiency and continuity of coding tasks. This advancement is significant for the AI/ML community as it aids in creating more context-aware and memory-efficient AI systems. The tool employs Qwen3-Embedding-8B for embedding user prompts and tool calls, which enhances search accuracy through a scoring system that combines cosine similarity with tool call relevance. The implementation maintains user privacy since everything is stored locally without telemetry. This revolutionary approach not only helps in extending the memory capabilities of AI agents but also allows configurations to adjust embedding weights, providing flexibility for varied use cases in AI-assisted software development.