Alma: an AI workspace built on a memory that survives every model

One AI workspace. One memory.

Most AI assistants forget. ChatGPT's memory holds a couple hundred short notes, and only inside ChatGPT. Claude and Gemini keep theirs in their own gardens. Switch models — or just open a new chat — and the context is gone. Re-pasting who you are, what you're building, and how you want your answers has quietly become part of the workflow.

Alma inverts that. The memory comes first; the rest of the product sits on top of it.

A memory that persists — and that you can read

It isn't a vector store stapled to a chat box. Three layers:

• Memories — discrete facts and preferences, pulled from conversations automatically. No tagging.
• Episodes — summaries of what happened, so long threads compress instead of bloating the context window.
• Procedures — repeatable patterns ("when asked for a draft, match the tone of the last three").

Retrieval is hybrid (keyword + semantic), scored so the relevant memory wins over the most recent one, and it runs on Cloudflare's edge — context assembly stays under 100ms in the common case. None of it is hidden: every stored fact shows up in a panel you can edit, delete, or export in full. The context belongs to the account, not the vendor.

Reachable from the tools already in use

curl -X POST https://alma.olivares.ai/api/v1/context/assemble \
  -H "X-API-Key: alma_key_..." \
  -H "Content-Type: application/json" \
  -d '{"user_message": "draft the launch email"}'
# → a system prompt with the relevant memories, identity and episodes

An MCP server brings that recall into Claude Desktop and Cursor. There's a documented REST API, a JavaScript SDK, and a VS Code extension. The top tier supports bring-your-own-keys.

A workspace, not a chat box — on one budget

Because the memory sits underneath, every surface uses it:

• Chat across Haiku, Sonnet and Opus (Opus 4.7 at a 1M-token window on Max).
• Image, video and music studios — Flux/Leonardo, Runway Gen-4, ElevenLabs.
• A calendar with two-way Google and CalDAV sync, fillable from chat.
• Documents, a Notion-style workspace, background agents, web search, writing tools.

One monthly budget covers all of it. No per-feature quotas, no separate video add-on, no image-credit wall. A picture generated in the studio inherits everything the memory already knows about the project it's for.

European edge, exportable on day one

Alma runs on Cloudflare's European edge. Data exports from the first session, and there's none of the regional gating that affects some assistants in the EU.

Pricing

Starter $14/mo, Pro $29, Max $99. Cancel anytime, 14-day refund. No data mining standing in for a fair price.

See it at olivares.ai — the API and MCP docs are on the developer page.