1. OKF Agent Memory Launches Git-Native Persistent Memory
OKF Agent Memory is a Git-native, domain-neutral persistent memory layer designed for AI coding agents. Built in Go with zero external dependencies, the MIT-licensed tool stores agent memories as plain Markdown files with YAML frontmatter directly inside the project's repository. It includes a built-in Model Context Protocol (MCP) server for seamless integration with tools like Claude Code and Cursor, utilizing local BM25 lexical indexing to eliminate vector embedding API costs and network latency.
- • OKF Agent Memory is a Git-native persistent memory layer for AI agents based on the Open Knowledge Format v0.2.
- • The system stores memory as plain Markdown files with YAML frontmatter directly inside the project repository.
- • It features a built-in Model Context Protocol (MCP) server for integration with Claude Code and Cursor.
- • It uses local BM25 lexical indexing, achieving search times under 300µs and eliminating vector embedding API costs.
- • The tool is written in Go with zero external dependencies and is released under the MIT License.
- • It employs progressive disclosure and a search-before-write principle to manage context bloat and prevent duplication.
Developers can equip their coding agents with fast, local, Git-native persistent memory without incurring vector database or embedding API costs.
2. Nous Research Adds One-Click Local Setup to Hermes Desktop
Nous Research has updated its Hermes Desktop application with a one-click local model setup feature that automates hardware detection, model selection, and runtime configuration. The free, MIT-licensed application manages its own llama.cpp build to support CUDA, Metal, Vulkan, and other backends. To prevent performance degradation, the software uses a color-coded system to show GPU memory compatibility, enforces a 4-bit quantization floor, and automatically unloads idle models after 15 minutes.
- • Nous Research added a one-click local model setup feature to its free, MIT-licensed Hermes Desktop application.
- • The software automatically detects hardware, selects a compatible model, downloads weights, and configures the llama.cpp runtime.
- • It supports CUDA, Metal, Vulkan, HIP, and CPU backends across macOS, Windows, and Linux.
- • A color-coded system indicates whether a model fits in GPU memory (green), spills to system RAM (amber), or is too large (red).
- • The application enforces a 4-bit quantization floor, hiding models that would spill into system RAM at 4-bit.
- • It automatically manages context windows up to 64K and unloads idle models after 15 minutes of inactivity.
Developers can quickly set up and run local models on their machines without manually configuring runtimes, quantization levels, or hardware backends.
3. Adaption Labs Launches Synthetic Dataset Generator
Adaption Labs has launched 'Invent a Dataset,' a new feature designed to generate synthetic training datasets directly from natural language task descriptions. By eliminating the need for a seed corpus, schema, or manual labeling guide, the tool simplifies the data preparation phase for fine-tuning. Developers can access the service via a Python SDK or REST API to generate prompt-completion pairs for supervised fine-tuning or chosen-rejected pairs for preference-based training.
- • Adaption Labs released 'Invent a Dataset' to generate training-ready datasets from task descriptions.
- • The tool does not require a seed corpus, schema, or labeling guide to initiate generation.
- • It is accessible via the Adaption app, Python SDK, and REST API, exporting to JSONL, JSON, CSV, or Parquet.
- • It supports prompt-completion pairs for supervised fine-tuning and chosen-rejected pairs for preference-based training.
- • Generated datasets integrate directly with Adaption's AutoScientist tool to co-optimize data and training recipes.
- • The service runs on Adaption's hosted platform and is billed based on expanded rows.
Developers can generate high-quality synthetic training datasets for fine-tuning without needing an existing seed corpus or manual labeling guides.