Why does Claude forget my house conventions?

Claude forgets your house conventions because conventions live in per-Project instructions (one Project at a time), per-repo CLAUDE.md files (Claude Code only, local to the machine), and per-chat corrections (gone at session end). There is no shared "house" object that every teammate's Claude reads from, so every contributor re-teaches the same rules. A memory layer pinned above Claude solves it.

Three things break convention memory.

1. Conventions are scoped to one Project or one repo. Claude.ai Projects keep their instructions to themselves. CLAUDE.md files live inside one repo. Move to a sibling repo or a sibling Project, and the conventions are gone. Your team has dozens of repos and a handful of Projects, so the rules fragment fast.

2. Auto memory does not leave the terminal. Claude Code's Auto memory writes notes about your preferences as it works, but those notes live in the local memory folder on the machine that ran Claude Code. They do not sync between contributors, machines, or to claude.ai in a browser.

3. Convention drift compounds over long chats. Even with conventions pasted into a Project, long chats summarize the rules. The model keeps the spirit ("use kebab case") and loses the specifics ("kebab case for folders, snake_case for migrations"). Subtle exceptions are the first to go.

The net effect: every teammate teaches Claude the same conventions, and the rules still drift mid-session.

Lost conventions look small in isolation and painful at scale:

• PR review becomes janitorial. Reviewers spend half their time fixing folder names and commit message format instead of catching real bugs.
• CI fails on style-only diffs. Migrations land in the wrong folder, lints catch banned imports, and pipelines break for reasons that should never have happened.
• Onboarding new teammates takes longer. New hires inherit chats that are still teaching Claude the same conventions you wrote down a year ago.

Anthropic offers two real but partial answers.

Project instructions and project knowledge in claude.ai let you pin conventions to one Project. You can also upload your team handbook or rules doc to project knowledge, and Claude reads it through Retrieval Augmented Generation when the content grows past the context window. Anthropic's own walkthrough of Projects and project knowledge is worth reading. The limit is the same as always: one Project, one account, no team-wide sharing without copy-paste.

CLAUDE.md files in Claude Code carry repo-specific rules and are read at the start of every session. They are excellent for the terminal. They are invisible to claude.ai, to teammates who do not use Claude Code, and to other AI tools your team uses.

These are good locally and weak at scale.

House conventions are a team artifact, not a per-Project preference. They apply to every repo, every chat, every teammate, every AI tool the team touches. Claude's memory model is built around per-Project, per-account scoping, which guarantees fragmentation the moment more than one person or one tool enters the picture.

The cleanest fix is to lift conventions one level above Claude, into a shared memory layer that every Claude session (and every Cursor, ChatGPT, or Gemini session) can read.

MemoryLake holds your house conventions as a shared, versioned source of truth.

• One shared rules Project for the whole team. Every teammate's Claude reads from the same MemoryLake Project. Update a rule once and the next chat anywhere uses it.
• Conventions survive long chats. MemoryLake retrieves only the rules relevant to the current file or task and re-injects them as the conversation grows, so summarization stops dropping specifics.
• Cross-tool convention enforcement. Cursor, Windsurf, Claude Code, and ChatGPT all read the same rules. No more contributor-by-contributor style policing in PRs.

MemoryLake scored 94.03% on the LoCoMo long-context benchmark, the top published result as of 2026, with millisecond retrieval and AES-256 end-to-end encryption.

1. Create a Project and load the house rules. Sign in to MemoryLake, open Project Management, click Create Project, and name it "House conventions". Upload your contributing guide, lint config, ADRs, and example PRs through the Document Drive, then add the highest-friction rules (folder structure, commit format, migration location) as named entries in the Memories tab.
2. Generate an MCP Server endpoint. Open the MCP Servers tab inside the project, click Add MCP Server, name it "Team Claude integration", and click Generate. Copy the API key ID, secret, and endpoint URL — the secret appears only once.
3. Connect Claude. In Claude Desktop or Claude Code, add the MCP server entry to the local config with the endpoint and Bearer token, then restart. For claude.ai in the browser, paste a single line in each Project's instructions that points at your MemoryLake project ID, and the REST API hydrates the rules per turn for every teammate.