Introducing Memory Prism 1
Today, we introduce Memory's first model, Memory Prism 1.
Based on open-source models, Memory Prism 1 is our first in-house compound model.
It delivers frontier-level performance (comparable to GPT-5.4 & Claude 4.6),
yet reduces cost by up to 85% compared to peer models.
Capability | Memory Prism 1 | GPT-5.4 xhigh | Claude Opus 4.6 | Gemini 3.1 Pro |
|---|---|---|---|---|
Tool Reasoning | 54.0 | 52.1 | 53.0 | 51.4 |
Deep Research | 92.5 | 78.6 | 91.3 | 81.9 |
Coding Agent | 58.6 | 57.7 | 53.4 | 54.2 |
Web Browsing | 83.2 | 82.7 | 83.7 | 85.9 |
Output Cost | $4.00 | $15.00 | $25.00 | $12.00 |
Better Answers, Lower Costs
Higher AI costs limit the user experience.
High costs force restricted usage,
making it hard to run long contexts, offload complex tasks, or use it widely across teams.
Memory Prism 1 was built to break this barrier.
Instead of relying on costly proprietary models,
we combined the strengths of open-source models into one system.
By hosting and running them ourselves, we optimized the cost-to-performance ratio for Memory.
As a result, you get frontier-level quality at a fraction of the cost.
What's Next
Memory Prism 1 is just the start.
We are fine-tuning next-gen models tailored specifically for Memory,
which will offer higher quality, lower cost, and more stable performance.
We are not trying to build just a larger model.
We want to make better answers
available to more people,
more frequently,
and at a lower cost.
Memory Prism 1 is the first step.