Are Blockchains the Perfect Vessels for Agentic Memory?
Continuing to believe in crypto x AI
Agentic memory is a massive driver in AI.
It has arguably taken us from the early days of toying with ChatGPT to the complete re-imagination of software engineering.
Example: Cursor just raised $900M for an app that ties together a VSCode extension + prompts and memory.
As traditional moats get weaker, memory may contribute to several moats all at once:
Switching costs. The more conversations you have with a memory-enabled app, the harder it will be to switch to another app even if it’s better. Context matters.
Improvement flywheels. Memories can be used by agents themselves to store learnings without getting explicit user feedback.
Network effects. This is more of a prediction but one I think we won't have to wait for long. Would you let a dating app ingest your ChatGPT memory to use it for matching? What about targeted advertising to find customers who may share a given pain point?
It also on the surface seems like something that could be implemented in a distributed system which may be good at retaining your data, letting you maintain ownership over it and providing composability with other applications that may want to consume parts of your context.
But to explore how memory could work in crypto, we need to understand how it works today.
AGENTIC MEMORY EXPLAINED
We are all familiar with the consequences of “enabling” memory in ChatGPT.
But what happens behind the scenes is more important:
Lets put ourselves in the shoes of an AI developer.
Suppose we are building an app that serves as the user’s personal meditation adviser.
The user will detail the ideal personality of their meditation adviser and then have daily audio-based meditation sessions followed by a reflective dialogue.
Memory categories
For the app to work well, the agent needs to remember:
Its personality that needs to be consistently applied from session to session and personalized for each user;
The user’s progression, development goals and any feedback on what they didn't enjoy (e.g., chanting meditations);
Details about the user such as personal experiences that could be addressed introspectively in future sessions.
I really like Mem0’s API so I’ll reference their docs frequently. Your mileage with other memory services may differ.
More generally, their API will remember user information across many more categories:
Short vs. long-term memory
Short-term memory is readily available context such as the current conversation window, time of day, etc. This context is typically computed in real-time and injected directly into the prompt and doesn't need to come from a database.
Long-term memory is information that needs to persist across sessions such as user preferences, facts about the users, tactics on how to better complete certain tasks and more.
Using long-term memory
The basic workflow is simple and shown below.
When dealing with a new query we need a “memory reader” to access relevant memories and at the end of our task we need to pass our response to a “memory writer” which can preserve long-term memories.
To use the API, you just need to provide the entire conversation window with Mem0:
From there it will figure out what to remember:
That’s it.
In future conversations, you can now use the API to access relevant memories.
While using memory APIs is simple, operating them at high performance is difficult, it requires:
Low latency (no more than 300ms for Mem0)
Providing good customization options without losing performance
Achieving high levels of accuracy (retrieving the right memories)
etc.
Mem0 uses both vector databases and graph databases depending on the structure of each particular memory concept.
It also provides memory management and expiry tools to manage overall memory footprint, reduce conflicts between memories, etc.
In short, using memory is simple, but building it isn't. This is a great recipe for creating enterprise value.
Which is why it would be pretty impactful for crypto to take a stab at this.
MEMORY AND CRYPTO
The challenge with trusting OpenAI or another platform with your “memory” data is two-fold:
You don't have ownership over potentially very sensitive personal facts and details that you wouldn't even share elsewhere. Many people already share medical details with ChatGPT to improve health care outcomes.
You lose portability of that information and can't reuse it for other applications.
Both map nicely to privacy tech and composability which are readily enabled with blockchains.
In the same way that I can port my wallet and assets from application to application, I'd love to do the same with agentic memory.
But many hurdles still remain.
It’s not clear that decentralized storage solutions are cheap enough to be used for agentic memory without altering the economics of agents.
And aggregators like OpenAI will compete by adding more and more value added services and we will all be tempted to keep growing our digital footprint there.
I do think crypto-native agents (for example, trading agents or even game NPCs) could provide a wedge to explore decentralized memory so I'm not giving up yet.