Stash / Blog · June 2026 · 7 min read

How to Make Claude Remember Things Across Conversations

Here's the situation you've run into: you close a Claude chat, open a new one, and Claude has no idea who you are, what you were working on, or what preferences you told it last Tuesday. Every conversation starts from zero.

This isn't a bug — it's how Claude is designed. Each conversation is isolated by default. But there are real workarounds, and some are much better than others depending on what you're trying to remember. This post walks through all of them honestly.

Option 1: Custom instructions

Claude.ai lets you set a system prompt that loads at the start of every conversation. Go to Settings → Custom instructions. Whatever you write there gets prepended to every chat.

Good for: Your name, your role, persistent preferences ("always respond in bullet points", "I'm a senior engineer, skip the basics"), your timezone, your communication style.

Bad for: Anything that changes. Lists, contacts, projects, tasks, records. Custom instructions are a static text field — you'd have to manually edit them every time something changes.

Verdict: Right for stable preferences that rarely change. Wrong for anything dynamic.

Option 2: Claude Projects

Claude.ai's Projects feature lets you attach files and instructions to a group of related conversations. If you're working on the same codebase or writing project across multiple chats, Projects keeps that context persistent within the project.

Good for: Long-running work projects where you need the same files and instructions every session. Team use cases where multiple people work on the same project.

Bad for: Anything outside of Claude.ai (API users don't get Projects). Dynamic lists or records that grow over time — you'd be editing an attached file, which is awkward. And every file you attach loads fully into context every time, regardless of what you actually need in that conversation.

Verdict: Useful for document-centric projects on Claude.ai. Not designed for lists, records, or cross-context memory.

Option 3: A shared notes file via a filesystem MCP

If you're comfortable adding MCP servers to Claude, you can connect a filesystem MCP that gives Claude access to a local file. You keep a notes.md that Claude can read and update.

Good for: Full flexibility. Claude can write anything, read anything. Works well if you're already technical and have the filesystem MCP set up.

Bad for: It loads the entire file on every read — no search. As your notes grow, the token cost grows too. A 2,000-word notes file costs roughly as much to load as a small article, every time. It also only works on one machine (your local filesystem).

Verdict: Works, but scales poorly. Fine for a few hundred words. Breaks down at a few thousand.

Option 4: A hosted MCP record store

The pattern that scales: a hosted service with a real database and MCP tools that let Claude search rather than load everything at once.

Instead of "give Claude a 3,000-word file," you do: "give Claude a database with 3,000 records, and a search tool that returns the 5 most relevant ones in 192 tokens."

This is what Stash does. You add one MCP connector URL to Claude, and you get:

• Collections — named lists for different types of records (contacts, projects, notes, context)
• Full-text search — Claude retrieves exactly what's relevant, not the whole database
• A context() tool — one call that loads your standing context (who you are, your current projects, your style preferences) from a reserved collection
• Works anywhere — any Claude conversation, any device, hosted not local

Which option is right for you?

The honest answer: for most Claude power users, the answer is both custom instructions and a record store. Custom instructions for who you are and how you work. A record store for everything that grows or changes.

How to set up the "Start my day" pattern with Stash

The most common use case after connecting Stash: a standing-context call that loads your morning brief. Here's exactly how it works.

1. Add Stash to Claude via the connector URL (takes 90 seconds — no code, just paste a URL)
2. Claude will use the context() tool to ask you to fill in your standing context on first run
3. Tell Claude your role, current projects, working style, anything you'd want loaded fresh each morning
4. From then on: open any new conversation, say "start my day" (or just let Claude call context() automatically) — your full context loads in one tool call

You can also add collections for specific things — a contacts list, an active projects tracker, meeting notes — and Claude can search any of them within the same conversation.

What about Claude's built-in memory?

Anthropic has announced memory features for Claude. As of June 2026, these are limited and not available to all users or via the API. If you're using the API — which most power users and teams are — you won't have access to those features.

The MCP connector approach works today, everywhere Claude works, including the API. It also gives you visibility and control: you can see what's stored, edit it, and delete it. Built-in memory is opaque by comparison.

The short version

Claude doesn't remember by default. The options are: custom instructions (static, good for preferences), Projects (good for document-centric work), a local notes file (flexible but scales badly), or a hosted MCP record store (good for anything that grows). For most people who've been frustrated by Claude's memory limitations, the record store approach is the one that actually holds up as usage scales.

Stash is free to start, takes 90 seconds to add, and requires no code.

Related: Claude Persistent Memory via MCP · The "Start My Day" Pattern for Claude · Claude Projects vs MCP: When to Use Each