Why does Claude forget my coding style?

Claude forgets your coding style because its persistent memory lives in two narrow places: Project instructions (capped, scoped to one Project, browser-only) and Claude Code's CLAUDE.md plus Auto memory (terminal-only, machine-local). Neither follows you between projects, accounts, or AI tools, so any style you teach in one chat evaporates the moment you switch context. A cross-model memory layer fixes that.

Three design choices in Claude's memory model explain the drift you see.

1. Project instructions are siloed and capped. When you create a Project in claude.ai and paste your coding rules into the instructions field, those rules apply only inside that one Project. Open a second Project for a different repo and Claude starts from a blank slate. The instructions field also has a soft character limit, so a full style guide gets truncated or summarized away.

2. Auto memory is Claude Code only. The newer Auto memory feature, available in Claude Code v2.1.59 and later, watches your terminal sessions and writes plain-markdown notes about your preferences. It is genuinely useful, but it does not exist in claude.ai, the iOS app, or the Chrome extension. If you write code in Cursor and review it in claude.ai, the style notes Claude Code saved never reach the browser.

3. Each new chat re-reads the same shrinking context. Claude Sonnet 4.6 and Opus 4.7 both ship with a 1M-token context window, which sounds infinite. In practice, Claude rereads Project knowledge through RAG when content exceeds the window, summarizes older turns, and prioritizes the most recent few hundred lines. Subtle style rules, like "wrap at 100 chars, never 80", are the first things the summarizer drops.

The result is a model that knows you exist but cannot reliably reproduce your house style across sessions.

Every regression to default style costs you minutes of cleanup and erodes trust in the output:

• Reviewers stop trusting AI-written PRs. When half your PRs use the team style and half use Claude's defaults, reviewers start nitpicking everything and merging slows down.
• Linter noise overwhelms real issues. Each new chat produces fresh lint errors that should never have appeared. Real bugs get buried under formatting churn.
• Onboarding new repos feels punishing. Every new Project needs the same 600-word style preamble. You either re-paste it or accept that Claude will fight your conventions for the first few turns.

Anthropic has shipped two features that partly cover this. Both have real edges.

Project instructions are the obvious answer for browser users. They persist across chats inside one Project, ride along with every conversation, and accept markdown. The catch: they live inside one Project, the field has a character cap that punishes long style guides, and there is no way to share instructions between Projects without copy-paste. Anthropic explains the setup in the official help article on Projects.

CLAUDE.md plus Auto memory is the Claude Code answer. You commit a CLAUDE.md to the repo with your rules, and Claude reads it at the start of every session. Auto memory then writes additional notes as Claude learns your corrections. This works beautifully inside the terminal, but it is invisible to claude.ai, to Cursor, to Gemini, and to any reviewer who opens Claude in a browser.

For one repo, on one machine, in one client, these are fine. For real cross-tool work, they leak.

The deeper problem is that coding style is a team artifact, not a per-Project preference. Your rules apply to every repo, every AI tool, every contributor, and every editor. Claude's memory is built around the opposite assumption: each Project is its own island, each client is its own silo, and each contributor's CLAUDE.md is their own file.

Add Cursor, Windsurf, ChatGPT, or Gemini to the workflow and the silos multiply. Style drift becomes the default state.

MemoryLake is a cross-model memory layer that gives your coding style one durable home, then feeds it back into Claude on every turn.

• One style guide, every Project. Store your house rules once in a MemoryLake Project. Claude reads them at the start of every chat, in every claude.ai Project, in Claude Code, and through the Chrome extension. No re-paste, no truncation, no character cap.
• Style survives summarization. MemoryLake's retrieval engine pulls only the rules relevant to the current file or language, then re-injects them as the conversation grows. Long sessions stop losing the thread.
• Same rules in Cursor, ChatGPT, Gemini. Switch to a different AI mid-PR and your style follows. The same memory powers code review in Claude and refactors in Cursor.

MemoryLake hit 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 your style guide. Sign in to MemoryLake, open Project Management, click Create Project, and name it something like "House coding style". Upload your style guide, linter config, and example PRs through the Document Drive (PDF, Word, Markdown, and code files all work), then add per-language rules as notes in the Memories tab so they ride along with every retrieval.
2. Generate an MCP Server endpoint. Open the MCP Servers tab inside your project, click Add MCP Server, name it "Claude integration", and click Generate. MemoryLake returns an API key ID, secret, and endpoint URL. Copy the secret straight away — it is shown only once.
3. Connect Claude. Claude Desktop speaks MCP natively, so paste the endpoint and Bearer token into your claude_desktop_config.json and restart. For claude.ai in the browser, drop a one-line system prompt into your Project instructions that references your MemoryLake project ID, and the REST API pulls the right style rules per turn.