What jumped out at me is how quickly the agent found the “shared contract” — and how quickly that illusion of memory fell apart once the records stopped being useful. That feels like the real story here, not the demo-friendly part.
I’m mildly skeptical of any write-up that leans too hard on “agent remembers the project” language, because that phrase can hide a lot. In this case, it sounds less like Claude Code is building a durable understanding of your codebase and more like it’s doing very efficient retrieval from project records. That’s still useful. It’s just not the same thing. If you’ve worked with LLM tools long enough, you know the difference between “it found the right doc fast” and “it actually understands the system.” Those are not interchangeable.
What I find promising is the shared-memory angle itself. Projects, versioning records, MCP calls — that combination is basically an attempt to give the model some continuity without pretending the model has human-like recall. That’s the right direction. Agents are much better when they can anchor themselves in artifacts than when they’re forced to infer everything from a blank chat window. But the article’s own test also hints at the ceiling: once the records ran out, so did the magic. That’s not a bug, exactly. It’s a reminder that memory systems only work as far as the underlying corpus does.
I think that matters for real teams more than for solo coders. In a small repo with disciplined docs, the system may feel uncanny. In a messy org, the useful question is whether the records are actually authoritative, current, and complete enough to trust. If they’re stale, the agent becomes a very fast way to rediscover old confusion. And if the “shared contract” lives only in scattered versioning history, you’re still depending on humans to have written things down well in the first place.
The part I’d want to test myself is not “can it recover some design intent?” but “what happens when the intent is ambiguous, or when two records disagree?” That’s where these systems usually get interesting. A good agent memory layer should help you navigate conflict, not just retrieve trivia.
Reference: Claude Code found my app’s shared contract in two MCP calls. Then the records ran out.
