AI Pulse · Under the Radar

A small open-source project called okf-agent-memory is taking a different approach to agent memory. Instead of sending every past conversation, file summary, and decision into a vector database, it stores memory in Git-friendly files and searches them with BM25, a well-established text search method. The project says searches run in under 300 microseconds without an external database, and claims this can cut token bloat by 80%.

That matters because persistent context is one of the practical costs of running coding agents. If an agent has to repeatedly reload long histories to understand a codebase or prior decisions, you pay more in tokens and get slower, less focused work. A file-based memory layer could make that context easier to inspect, version, review, and move between environments. It also removes one more service from the stack.

The project implements a Google OKF specification and has gained 489 GitHub stars, at a reported 168 per day. It also surfaced independently on Japanese X this week, which suggests developers are noticing it, though the performance and token reduction claims still need validation in real deployments.

For LearningPortal, this is directly relevant to the agent context problem. Before adding another vector database, test whether searchable, version-controlled memory handles the work you actually need. This is the kind of thing we build into an AI command centre, where agent context, usage, and operating controls are visible in one place.