for sure! we're seeing this work well for many of our users
Ontologies become rigid when agents can only change the data, not the schema
In @omnigraphdb graph config is 100% declarative, so background agents can propose changes to the schema file when their understanding of the domain evolves
You can keep the schema and config in GitHub, review changes as PRs and deploy them with CI.
This way domain model can evolve without you giving up control.
Happy to give you a quick demo
Agent memory is the wrong abstraction.
Memory is a human concept.
Our memory is lossy, emotion-driven and embodied.
Which doesn't make much sense for current AI.
Agents are modular and based on stateless LLMs.
They need explicit concepts, relationships, and ground rules: an ontology they can reason with.
Once they have it - they get insanely clever with retrieval and signal detection.
That's what a context layer provides (or should provide, at minimum)
And context layer is the right abstraction btw
Agents don't need to forget
We (humans) store complex, multi-sensory holographic 5D memories and we have limited capacity for those.
Storage is super-expensive so we must prune it.
Agents can keep all context forever, with full lineage and provenance.
On object storage it's quite cheap.
What agents need is the right ontology optimized to keep things current and deconflicted.
For example:
superseded decisions and policies clearly labeled as so, with immutable history
causal chain reasoning proven wrong turned into learnings, with pointers to immutable case records
and so on
Agents are trained on explicit human artifacts
and ontologies are at the core of them
Memory doesn't produce stable artifacts, but rather
ephemeral "thoughts" and unstable mental states
Those are biological.
The artifacts we produce are cultural:
atomic and stable (books, articles, research papers, db records, ...)
define and enforce invariants
clean & curate ruthlessly, fight the bloat
code is the source of truth
prove all automations on evals first
make all specs 2 parts: [for humans | for agents]
→ delete [for agents] part after implementation
cognitive resource > tokens
prompt for the design, not the patch
@Killian_Kirk Agents learn with time and I trust them
I do not use confidence levels for my personal graph. Overall they are useful, but require too much effort to setup properly
I have to admit: I'm enjoying Studio way too much
Studio [our graph IDE for @omnigraphdb]
built on GPUI: Rust-native UI framework from @zeddotdev
Feels sooo snappy compared to React
🦀🦀🦀
You can try it right now↓
"dreaming" sounds misleading to me too
Human dreaming is kind async memory processing, but again, it's unpredictable and non-codified
I run a bunch of agents on my graph nightly:
enrichment, deconfliction, connection discovery, verification and so on.
All clearly defined and predictable, with full provenance and control
They write to branches which another agent or I merge based on policies
@Sam4Jef both actually
I use @omnigraphdb with fully declarative config and branches.
I give agents incentive to propose ontology changes when they see branch data do not fit.
Notion's biggest problem - it's a bundle.
And we're living through the greatest unbundling in software history:
moving from a few mediocre omni-apps
to an abundance of awesome lego blocks
Notion will have to deal with this new reality eventually
But for now, repositioning as the company brain seems smart
"So explanation does, after all, play the same paramount role in pure mathematics as it does in science. Explaining and understanding the world – the physical world and the world of mathematical abstractions – is in both cases the object of the exercise. Proof and observation are merely means by which we check our explanations." The Fabric of Reality.
2 questions come to mind:
Does a proof also contain elements of explanation?
In a sense, each step "explains" the next, so it's more than just mechanical verification.
Is math getting closer to biology?
Like we have the genome but still need to figure out the organism.
Will more conjectures come from decoding the AI output?
@mckaywrigley the math is pretty much solved
now pick a practical algo:
indexing, query planning, data partitioning,
entity resolution, compression, caching, ranking, anomaly detection
...
figure out the application
profit🤷♂️