memgraph v1.9.1 ships a small but consequential change: a content discipline section now lives inside the guide, and a 4-question gate runs before every memgraph remember. The headline is "memorize what code cannot tell you" — if the answer to a question is already in the source, the repo, or the commit log, it doesn't belong in memory.
The motivation is straightforward. Memory stores that duplicate what the codebase already says get noisy fast, and noisy memory is worse than no memory because agents trust it. This release pushes the filter upstream: don't write it down if you could just read it from the repo.
What changed
- Content discipline section in the guide. Both the JSON and human-readable guide output now include a content discipline section, so any agent reading the guide picks up the rule set without an extra fetch.
-
4-question gate before every remember. Before
memgraph remembercommits anything, four questions check whether the content is something code itself can tell you. If it is, the memory is rejected. This keeps the memory store focused on non-obvious knowledge — decisions, caveats, deployment quirks — rather than duplicating what's already in the repo. - In-repo skill. The full skill lives at the path below, versioned with the repo and discoverable by other agents working in the same codebase:
.agents/skills/memgraph-content-discipline/SKILL.md
- GitHub link in guide output. The guide now links to the full skill on GitHub, so a human reading the guide output can jump straight to the source.
- .gitignore fix. A previous .gitignore rule was matching the skill directory as binary, which would have excluded it from the repo. Fixed.
The diff is 4 files changed, +152/-1. Small release, but it changes the default behavior of every remember call going forward — the gate runs unless you explicitly bypass it.
Top comments (0)