It remembers.
What your users say becomes memory your AI can find again next week, next month, next year. Nobody has to repeat themselves.
Long-term memory for AI agents
MindDB gives your AI a memory that lasts. ItMindDB remembers what your users say, knows what changed, and hands your model one short memory page with exactly what matters.
These examples are illustrations, not recorded answers.
What it does
What your users say becomes memory your AI can find again next week, next month, next year. Nobody has to repeat themselves.
People move, plans shift, records get beaten. When a fact changes, the new one takes its place and the old one becomes history, so your AI answers with what's true now.
Some answers take more than one step. MindDB links one memory to the next until it reaches the one your question needs.
See it work
A user mentions something in March. In June, in a brand-new conversation, they ask a question that depends on it. Switch the memory on and off, and watch the same model answer.
Illustration. Both replies were written for this page.
March 3
My telescope finally gave out. I want a new one before my birthday in June.
That's a shame. Tell me if you'd like help choosing one.
Three months later · a new conversation
Any ideas for my birthday?
Memory page · handed to the model
Happy to help! What kinds of things are you into?
Your telescope gave out in March, and you wanted a new one before June. Want me to shortlist a few?
Nothing to go on. The user has to start over.
One memory page. The right answer.
One memory page, not the whole history
Most AI remembers by dragging the whole conversation along, so the cost grows with every turn. MindDB hands your model one short memory page, however long the history.
We ran the same 1,000-message conversations through five ways of remembering, all on one model. MindDB matched the most accurate and used the fewest tokens by far.
Same conversations. Same questions.
MindDB
93 of 100 right
1× tokens
Everyone else spent more
Mem0
2× tokens
Summarize and compact
5× tokens
Letta
10× tokens
Sliding window
12× tokens
2–12×the tokens, everywhere else.
No one was more accurate.
A pre-registered study: 12 synthetic conversations of 1,000 messages each, with questions asked at six points along the way and scored by a fixed rule. Every system used the same local 27B open-weight model. The rings show answers right out of 100. "×" is input tokens per right answer, relative to MindDB.
| System | Right | Tokens per right answer | Against MindDB |
|---|---|---|---|
| MindDB | 93.2% | 64,685 | 1× |
| Mem0 | 84.7% | 124,950 | 1.9× |
| Summarize and compact | 83.6% | 335,274 | 5.2× |
| Letta | 92.4% | 645,587 | 10.0× |
| Sliding window | 70.0% | 769,474 | 11.9× |
Mem0 is Mem0 OSS 2.2.1, set up as its README's chat loop. Letta is Letta 0.16.8 with its default agent and a 65,536-token window; it ran 6 of the 12 conversations, so its row is descriptive. In two of those six, our local server cut prompts that had outgrown the window. Neither system is tuned for this model, and results on other models may differ.
Of the accuracy gaps, the registered tests confirmed the one over the sliding window. The others, including a one-point gap to Letta that we call a tie, are reported as measured. Letta and Mem0 are the names of their owners' products; we are not affiliated with either.
Bring your own model
A memory page is plain text, so your AI's memory is never tied to one vendor or one version. Start on a hosted model, move to an open one, run it on your own hardware. What your AI knows comes with you.
Our benchmarks so far ran on local open-weight models.
What you can build
Every order, fault and promise is there the moment the next conversation starts.
Preferences, plans and the people in someone's life stay current as they change.
Decisions and their reasons carry across days, without dragging the whole log along.
Companions and game characters remember each player, and never mix them up.
Sources, findings and open questions add up over months of a project.
Your platform's agents get a memory under your name, on your hardware.
Early access
We're opening MindDB to a small number of teams, and licensing it to companies that need memory on their own hardware. Tell us what you're building and we'll be in touch.