Mainplane Gives Coding Agents Searchable Memory to Beat Context Limits

Mainplane is a new open-source agent harness that stores complete session transcripts and lets coding agents retrieve earlier details after a context window fills. The MIT-licensed project, written in Go, targets a costly failure mode in long-running coding sessions: compaction can delay the handoff and discard exact commands, paths, errors, and measurements.

The project also provides a self-hosted control plane for connecting models, worker devices, shared storage, and messaging channels. Developers can run it on one computer or coordinate agents across several machines without committing to a single model provider.

Why compaction gets expensive

Language models can process only a finite amount of text in each request, known as the context window. Once a coding session approaches that limit, agent tools commonly ask the model to reread the conversation, produce a condensed summary, and initialize a fresh session from that summary.

Mainplane’s critique focuses on the latency, token cost, and information loss introduced by this handoff. In a benchmark post, the project’s author reported results from a roughly 400,000-token session using a model identified as Opus 5.5:

A prompt cache allows a model provider to reuse previously processed input, reducing the cost and latency of repeated context. An expired cache forces the service to process the session again. The figures above come from one author-run comparison and have not been independently reproduced across providers, models, or workloads.