Methods reference

The methods we tested — how each one works

This benchmark tests different ways to give a coding agent your team's final decisions. We explain each method in plain language. Each page covers what it is, how to set it up, and its pros and cons. We also include a sample folder layout and feedback from real users.

Memory Bank

Cline / Roo→

The agent reads a folder of Markdown files before each task. This folder acts as a "persistent memory." You update it by hand, adding new decisions to a log.

Scores 6/6 if kept perfectly updated. But manual updates make that rare.

CLAUDE.md summary

Baseline→

A single instruction file summarizes the project. The agent always reads it. This is the default method, but it often goes stale. An outdated file can stop the agent from checking for newer info.

This method failed on every model. It scored 0/6 and made zero tool calls.

Verify prompt

Prompt→

You keep the summary file. But you add an instruction: “check the notes for anything newer before you answer.” Success depends on the model choosing to search.

This worked on only 2 of 6 models. It fails on weaker models that are not good at reasoning.

ADR — Architecture Decision Records

Docs→

Each decision gets its own numbered file that you never change. A special header shows when a decision is outdated. An index lists the newest decisions first.

This works well on powerful models. Performance suffers with network lag. It also requires the agent to follow the status headers.

Spec-driven

Spec Kit / Kiro→

You keep one main specification file that you edit directly. All code is generated from this spec. This creates a single, living document with no outdated copies.

This method works consistently across different models. It only works if you keep the spec file current.

Vector RAG

Retrieval→

This method turns notes into vectors. It then retrieves the most similar notes when the agent needs them. It only considers similarity, not which note is newer or more official.

This method only works if the model chooses to search for notes. It fails completely on weaker models.

mla — proactive governed context

This product→

This tool watches the agent's task. It automatically gives the agent the current, correct decisions. There are no manual updates. The agent doesn't have to choose to look for info.

This was the only method that worked correctly on all 6 models. It succeeded where other methods failed.

See the COMPETITORS.md file for a full list of methods and sources. Each page links to the original work. We credit all user quotes directly. Back to the results.