OKF Agent Memory Released, Cuts Tokens by 80% Without Vector DB

A new tool OKF Agent Memory following Google's Open Knowledge Format(OKF) v0.2 specification has been released. This tool manages the long-term memory of AI coding agents by utilizing Markdown files (including YAML front matter) in the knowledge/ directory within a Git repository, without a vector database.

Core Mechanism and Tech Stack

Existing methods like CLAUDE.md or AGENTS.md were inefficient because the entire file content was included in the prompt for every request as the file grew. OKF Agent Memory applies the Progressive Disclosure approach, retrieving only concept files relevant to the query and injecting them into the prompt.

• Tech Stack: A single executable written in pure Go language with no external dependencies.
• Search Engine: Performs in-memory BM25 score-based search without external embedding APIs.
• Integration: Acts as a Model Context Protocol(MCP) server, allowing connection with Claude Code, Cursor, etc.
• Trust Classification: Formalizes generated (speculation) and verified (human-confirmed) grades via YAML front matter to manage information reliability.

Performance Benchmark Results

The method of inputting the entire document was compared with the method of retrieving and inputting only relevant concepts across 9 models.

• Token Reduction: Input tokens decreased by approximately 80% (3,034 → 603).
• Latency: While varying by model, time to first token was reduced by 1.1x (gpt-5.6-sol) to 5.2x (ministral-3-14b).
• Local Inference: For local models (prism-ml/bonsai-27b, etc.), 3.8~4.7x acceleration was confirmed due to the reduction in prompt pre-computation.
• Policy Compliance: Both methods passed compliance checks for 4 items, including encryption policies.

Limitations and Installation

Currently, Korean search is not supported. The tokenize logic in pkg/okf/search.go only processes English and numbers, so no results are returned for Korean queries. Installation involves generating the bin/okf executable via make build, with support for GitHub, the homepage, and Homebrew. The license is MIT.