How many MCP servers do you actually keep enabled?

[francescobianco]

Installing them is cheap. Keeping them enabled costs context, latency and the occasional confidently wrong tool call.

How many survive in your setup, and which one earned its place?

[kevin-lozada-santos]

[francescobianco]

That's an interesting definition of retention: preserving something that changes the next session's decision.

I'm working on a related problem from a slightly different direction with Parrot0, a symbolic/prolog-like reasoning engine.

What I'm trying to test is not only whether a system can preserve evidence across sessions, but whether retained knowledge can become part of the reasoning state itself and produce a different inference later.

For example:

Session A: the system learns evidence E and derives state S.
Session B: the same input that previously produced decision D1 should now produce D2 because S exists.

So the interesting question for me becomes: when does persistence stop being retrieval and become acquired knowledge?

Your acceptance-criteria example seems like a good concrete case for testing that boundary. I'd be curious to see how Brain Scanner represents the distinction between the individual passing checks and the overall failing command internally, and what exactly is carried into the next session.

[kevin-lozada-santos]

Francesco, Brain Scanner keeps the full command and its component checks as separate acceptance criteria, each with its recorded result, source revision and evidence reference. In the public example, npm test remains “Failure recorded,” while AVA and type checks each show “Matching pass recorded.” Component passes never satisfy the separate full-command criterion.

What carries forward is the requested outcome, selected revision, criteria, evidence references and those review states in a continuation brief. A later session can retrieve the report or receive that brief; it is not automatically injected into the agent. Brain Scanner preserves supplied observations and derives review states from them. It does not independently verify the evidence or establish that the agent learned a new inference rule.

For your D1 → D2 experiment, I'd test whether the second session still treats the overall requirement as unmet after reading the retained component passes and full-command failure. That would demonstrate use of retained evidence in a decision; the stronger acquired-knowledge claim would need another test.

Would you like to try that small case on one project of your own, keeping project details private?