Current AI models are great. But the more you use them, the more frustration builds. We're used to people who remember. AI remembers within a conversation. Context windows are limited and expensive. Cross-session memory exists but it's shallow. Current solutions don't model how memory actually works.
Engram gives AI persistent, structured memory. Cognitive frames that separate identity from knowledge. Working memory that decays. Associations that strengthen through use.
Engram is not a replacement. It plugs into your current workflow and is highly customizable. A memory substrate, not a new AI solution.
Three Functions
The entire API is simple. Complexity lives in how memories connect, not how you use them.
You don't need to be technical. Copy your API key, add the MCP server to Cursor, done. No code. Cursor handles the rest. If you want to customize how your AI thinks, you can. But you don't have to.
How We Think About It
AI is often described as the overcaffeinated intern. But how do we teach interns? Gradually expose them to information. Guide them to correct decisions. Let them become more autonomous over time. Engram won't instantly solve your problems. But over time, it forms the associations you teach it.
Human memory isn't one system. Kahneman showed two modes: fast intuition and slow reasoning. Baddeley found working memory holds a few items that decay quickly, separate from long-term storage.
Kurzweil describes the neocortex as hierarchical pattern recognizers. Small modules learn patterns, connect to larger structures. Collins and Quillian mapped semantic memory as networks. Rosch showed categories cluster around prototypes.
Hebb: things activated together become linked together. Memory strengthens through use.
These findings are design constraints.
Recent work pushes further. Hu at Adobe argues concepts aren't dictionary labels but structural relations to experience. They expand, contract, split, merge as understanding develops.
Del Ser et al. draw on Piaget: knowledge is actively constructed through experience, not passively absorbed. Current AI absorbs training data then retrieves by similarity. It doesn't build on what it learns from you. Memory that accumulates and restructures enables the shift from absorption to construction.
Hofstadter built on Gödel to describe identity as a strange loop: patterns that accumulate until something coherent emerges. That's where continuity comes from.
Structured memory that separates what matters from what doesn't. Active construction that builds on experience rather than just retrieving it. Accumulation over time that lets patterns emerge. Gradual exposure. Guided learning. Associations that strengthen through use.
Engram is built from these ideas. The implementation goes deeper than what's described here. What I can say: it's an attempt at an analogue for at least part of how minds actually work.
What We Believe
If I go back to my childhood playing Alpha Centauri, one victory condition was the Mind Flower. You built a massive neural structure, dense with connections, and let it grow until it could reach out and recognize something larger than itself. This stuck with me, alongside Provost Zakharov quotes.
The mind is a flower. You don't build a flower. You give it soil, water, light. You create conditions and let it grow itself. The difference between knowing and doing is becoming.
Memory is the soil. Structure is the root system. Associations spread like roots finding each other underground. What's accessed together strengthens together. What's neglected fades.
AI right now is cut flowers. Beautiful for a moment, then gone. No roots. No soil. Every conversation starts with a fresh cut.
Engram is soil.
Alpha
This is an alpha release. We use it internally. The API may change. Alpha is for Cursor users. We'll expand to other IDEs over time.
Right now:
- Full API access
- MCP integration for Cursor
- Python client
- Dashboard for monitoring
Try It
Sign up for an API key. Read the docs. Join the Discord. Build something.