Read-only environment probing moves a 40-question agent memory benchmark by 1.2 questions

Three points of pass rate on 40 questions is 1.2 questions. Probing added those three points in the CLBench run described in MemTensor's post: 73% mean pass rate with memory plus environment probing, against 70% for memory alone. The no-memory baseline was 39%, so memory by itself bought 31 points, roughly ten times what probing added on top.

The cost side reads better. Average queries per question fell from 5.6 to 4.7, and task-agent cost fell from $1.99 to $1.68, about 16% off each. Per the post, those figures exclude the separate distillation and curation stages. Across the 40 questions, the task-agent saving totals $12.40.

For the three points to transfer, your environment has to move the way this benchmark's does. The run used GPT-5.4 in a GitHub Copilot SDK harness on questions involving schema changes, and probing is a way of catching that drift after the task is over. Static schemas give the curator nothing to find.

What the curator does is narrow. It runs after a task ends, reads the environment without writing to it, and revises or skips a record before any later task can use it. Read-only is what keeps the check cheap: a coding agent inspects a symbol definition, a database agent reads a schema or queries a limited sample.

MemTensor's worked example is a database agent that finds the records it needs in a table called customers_current and saves the note "Use customers_current for active accounts." The query may have worked for one region or one reporting period, and the note leaves those conditions out. Reading the table definition and the relevant business rules would establish where the advice holds. A month later, the post notes, the table might be replaced by a compatibility view that updates less frequently.

The other failures the post lists have the same shape: an API workaround still recommended after the fix, a temporary approval process saved as a permanent procedure, a failed command recorded as a successful solution. Retrieval works in all of them. "Retrieval can find the note, but the agent still needs enough information to judge whether it applies," the post said.

Sizing the check to what is being saved is the part I would copy. A preference about report length needs less review than a remembered procedure that could change access permissions or authorize a payment. Procedures with side effects cannot be settled read-only at all; they need a suitable test environment or other evidence of a successful result. MemOS itself ships operations for adding, finding, correcting and removing memories, and MemReader turns conversations, documents and images into memory items carrying source information.

Choosing the authoritative source and writing the environment checks stays with the application. The post also argues for keeping the original conversation and tool results as evidence, so a team can inspect how a summary was produced when something goes wrong.