Your agents keep forgetting, and your token bill pays for it

Open a fresh session with any AI agent and watch what it does first. It re-reads. The same project background, the same file structure, the same "here is what we decided last time," pulled back into the context window from zero. It has no memory of yesterday's work, so it rebuilds the whole world before it can help you. And you pay for every one of those tokens, every single time.

That cost is quiet because it hides inside what looks like normal usage. A long-running project can burn more tokens re-establishing context than it spends on the actual task. Multiply that by every session, every teammate, and every connected tool, and the bill climbs for a reason that has nothing to do with the work getting done.

The bottleneck was never the agents

We (Nils and Florian) spent a long time assuming the models were the constraint. They are not. The models are good. What breaks is everything around them: the connection to your tools, and the context that evaporates the moment a session ends. An agent that cannot remember is an agent you have to re-brief forever.

So the real problem is a memory problem. If context survived between sessions, the agent would stop paying the re-reading tax, and it would stop asking you the same questions it asked last week.

A memory layer next to your MCP servers

That is what we have been building. Tulimoa Memory is persistent memory for AI agents, served as its own MCP server next to the tools your agent already uses. It holds context across sessions with two simple moves: remember to store what matters, and recall to bring it back. Once a night it consolidates what was stored, merging duplicates and turning details into durable knowledge. The agent stops rebuilding the world from scratch, so it stops re-reading (and re-paying) for context it already had.

This is the brand core, not a side feature. The memory layer that shrinks your token bill.

A few concrete things that are live today:

• One MCP endpoint, memory.tulimoa.com/mcp, that works in any MCP client next to your other servers, with one key.
• A morning briefing that says what was consolidated, reinforced and forgotten, and what is still open.
• Privacy built in: email addresses, phone numbers and IBANs become placeholders before anything is stored, and a person can be erased by crypto-shredding.
• The database sits in the EU (Frankfurt), with a dashboard in English or German.

Honestly, it is early

We are not going to dress this up. Tulimoa Memory is young. There are rough edges, and we are shipping fixes constantly. There is no plan and no subscription: every call costs $0.002, and you get a $1 welcome credit once to start.

If your agents keep forgetting, and your token bill keeps climbing for context you have already paid to load, come try Tulimoa Memory. Create a memory in the dashboard, connect it to your agent, remember something, close the session, open a fresh one, and recall it. That loop is the whole idea, and it is the part we most want your feedback on.