Anthropic has added memory to Claude.ai. Claude can now remember things about you across conversations — your name, preferences, ongoing projects — without you repeating yourself every session.
That's genuinely useful. So where does MCP-based memory like Stash fit in?
Short answer: they solve different problems, and for power users, you need both.
When Claude's native memory is enabled, Claude automatically saves facts about you as it learns them. Over time, it builds a profile: your job, your communication style, recurring topics, things you've mentioned caring about.
This works well for:
Claude decides what to save. It's automatic, ambient, low-friction. You don't have to do anything.
Native memory is passive and unstructured. Claude picks what it thinks is worth saving. For heavy users, this creates real gaps.
You can't ask Claude "show me everything you've saved about my clients" or "what projects have I mentioned this month?" The memory is there, but you can't interact with it as data.
If you use the Claude API (for integrations, automations, custom tools), native memory doesn't follow you. API calls start fresh. Each API session has no access to what Claude.ai remembered.
Native memory is prose — facts and notes, not structured records. You can't store a collection of 500 contacts with fields, search by tag, or retrieve "all notes on Project X." It's more like a memory journal than a database.
For users with substantial records — customer lists, research notes, decision logs — the unstructured native memory can't scale. There's no pagination, no search, no capacity for thousands of records.
You don't control the data. You can view and delete memories, but you can't export your full memory as a structured file you own and can move.
MCP-based memory (like Stash) is intentional and structured. You tell Claude what to store, in what shape, in collections you define. Then you search it with full-text queries, and Claude loads only what's relevant — not the whole database.
| Feature | Native memory (Claude.ai) | MCP memory (Stash) |
|---|---|---|
| Setup required | None — automatic | 30 seconds (connector URL) |
| What gets saved | Claude decides | You decide |
| Structure | Prose notes | Named fields, collections |
| Search | Not queryable | Full-text search |
| Works with Claude API | No | Yes |
| Record capacity | Limited (ambient facts) | Thousands of records |
| Data ownership | Anthropic's servers | Your data, your account |
| Best for | Personal preferences | Work records, lists, notes |
Native memory is excellent for: who you are, how you like to work, your communication style. It's zero-effort context that makes Claude more useful in general conversation.
MCP memory is excellent for: structured data you need Claude to retrieve — client records, project notes, a reading list, a decisions log, daily context for "start my day." It's intentional storage you control.
They don't compete. They layer. Let native memory handle the ambient background. Use Stash for the records you actually need to find again.
Stash is a hosted MCP server. You connect it to Claude in about 30 seconds:
context() to see your standing context, or start adding recordsThe free tier includes 2,500 records and 50 searches per month — enough to evaluate whether structured MCP memory fits your workflow.
Add Stash to Claude
Structured, searchable memory that works with any Claude session — including API calls.
Connect Stash →