Gemini Memory Problems: Why Google's AI Forgets You

Updated January 2026 | 6 min read

You tell Gemini you're a freelance consultant who bills hourly. Three conversations later, it suggests you "implement company-wide policies." You explain your tech stack includes Obsidian and Claude Code. Next session, it recommends Notion and ChatGPT.

Google just rolled out Personal Intelligence—a memory upgrade that connects Gemini to your Gmail, Photos, and other Google apps. Sounds good on paper. In practice, it's amnesia with a search function.

How Gemini's Memory Actually Works

As of January 2026, Gemini has two memory systems that don't talk to each other:

Personal Intelligence pulls data from connected Google apps. You ask about a meeting, it checks Gmail. You need a photo, it scans Google Photos. It's not remembering—it's searching your account in real-time.

Core Memory uses vector embedding to save "important user facts" across conversations. You can't disable it, only opt out of specific data sources or delete individual memories. Google says this builds a "personal memory bank," but it's selective and opaque about what it keeps.

The one-million-token context window (two million for Gemini 1.5 Pro) sounds massive. That's about 750,000 words of working memory per session. But it resets. Each new conversation starts blank unless you manually reference previous threads or wait for Personal Intelligence to surface something from your app data.

Google's roadmap includes context fusion (auto-retrieving past chats), adaptive compression (summarizing earlier context), and cross-app recall (remembering goals across Gmail, Docs, and Sheets). None of that exists yet. Right now, Gemini forgets everything outside your current session and your Google ecosystem.

What Gemini Gets Wrong About Memory

The architecture assumes your life lives in Google apps. If you use Obsidian for notes, Slack for team communication, or any non-Google tool, Gemini has no access. Personal Intelligence is turned off by default, and even when enabled, it's limited to what Google owns.

Core Memory saves facts, but you don't control what counts as "important." You tell Gemini you work in real estate. Sometimes it remembers. Sometimes it asks again. The system learns patterns but won't show you the full list of what it's stored. You get a "memory updated" signal, but not transparency.

The bigger issue: retrieval vs. retention. Personal Intelligence doesn't persist your preferences across tools. It reads your Gmail to answer questions about a client, but it won't remember that client's communication style for next time. It references data, doesn't absorb it.

Session isolation means every conversation is a cold start. You spend ten minutes explaining your workflow. Gemini drafts something useful. You close the tab. Tomorrow, you're explaining the workflow again.

The Workarounds (And Why They Fall Short)

Connect all your Google apps. You enable Personal Intelligence, link Gmail, Calendar, Photos. Now Gemini can pull info from those sources—but only when you ask questions it can search for. It won't proactively apply context. And if you work outside Google's walled garden, you're stuck.

Manually paste context every session. You keep a doc with your bio, preferences, current projects. Each conversation, you copy-paste the whole thing. It works until you forget, or until the doc gets too long and you're wasting 200 tokens on setup.

Use the same thread forever. Some users never start a new conversation. They keep one mega-thread so Gemini retains context. It breaks down when the conversation history gets so long that relevant context gets buried. And you still lose anything outside the current thread.

Write detailed prompts every time. "I'm Victor, a freelance consultant who works with real estate databases and SEO clients. I use Obsidian for notes and Claude Code for AI workflows. Here's what I need..." Every. Single. Time. It's exhausting, and half the details still don't stick.

None of these fix the core problem: Gemini doesn't have a single, persistent source of truth about who you are.

The Alternative

Google's approach treats memory as retrieval—search your apps, find the answer. That works for factual queries. It fails for identity.

The file-based approach flips it: one markdown file contains everything. Your role, your tools, your clients, your voice, your context. CLAUDE.md lives in your Obsidian vault. Claude Code reads it every session. No searching, no guessing, no resets.

You tell Gemini you're a consultant once, maybe it remembers. You tell CLAUDE.md once, it's permanent. Every AI session starts with full context. No setup, no copy-paste, no explaining yourself again.

It's not locked to one vendor's ecosystem. Obsidian syncs across devices. CLAUDE.md works with Claude Code today, adapts to other tools tomorrow. You own the file, you control the memory.

Gemini's Personal Intelligence requires you to live inside Google's infrastructure. CLAUDE.md lets you use whatever tools you want and bring the AI to your workflow, not the other way around.

Google's betting on cross-app recall and context fusion as future features. File-based memory already does this. Your vault connects your notes, your projects, your daily logs. The AI reads all of it, not just what it can search.