AI Researcher & Engineer
Amsterdam
Joined May 2010
- Tweets133
- Following254
- Followers92
- Likes328
Our paper on neuro-symbolic meta-policies for knowledge-graph memory was accepted at ISWC 2026, with @VinFL and @michaelcochez
Don't hardcode how an agent answers, explores, and forgets. Let it learn which symbolic rule to apply, while execution stays symbolic and inspectable.
Memory is an RDF graph where every fact carries when it was added, last used, and how often it was recalled. The named rules rank over those values, and a learned controller picks one per step.
Project page: humem.ai/projects/roomkg-met…
Paper: arxiv.org/abs/2607.18368
Code: github.com/humemai/roomkg-me…
Our paper got accepted at #ROMAN2026 🎉
The idea: a robot shouldn't start every new collaboration with a blank memory. We let it reuse a pattern from past human-robot teamwork, stored as a knowledge graph, before the task even starts.
In our study this took rescue success from 26% to 41% and cut task time by ~5 min, mostly early on.
With Emma van Zoelen and Mark Neerincx (TNO, TU Delft).
Project: humem.ai/projects/co-learnin…
Paper: arxiv.org/abs/2606.18836
We are happy to share that our paper, "Short-Term-to-Long-Term Memory Transfer for Knowledge Graphs under Partial Observability", has been accepted at RLC 2026.
This work was done together with @VinFL and @michaelcochez
By making memory transfer explicit and learnable as a keep/drop decision over knowledge-graph facts, we move a step closer to AI systems whose memory is not only effective, but also interpretable.
I’ve been working on embedded Python bindings for ArcadeDB so I can use it directly from Python without a driver hop.
What I like about this setup is that it keeps local experimentation simple: SQL, OpenCypher, and vector search from Python,
and the package ships with a lightweight JVM and the required JARs, so there’s no separate Java setup to deal with. The wheel is also much smaller now, around 74MB, and it supports multiple platforms.
This has been a nice fit for the kind of memory-system and retrieval work I care about.
Learn more at taewoon.kim/2026/01/28/build…
Everyone says “knowledge graph,” but few define it precisely.
This post walks through the formal foundations of knowledge graphs, unpacks Hogan et al. (2020)’s definition, and compares how RDF triple stores and property graph databases like Neo4j actually implement them.
You’ll learn why industry favors property graphs for AI/ML pipelines (fast traversals, flexible attributes, embedding integration) while RDF still anchors interoperability and reasoning.
Read the full guide → taewoon.kim/2025-10-06-knowl…
🧠 From REINFORCE to PPO: The On-Policy Algorithms Behind Modern LLM Fine-Tuning
Large Language Models like ChatGPT don’t just learn from text — they’re fine-tuned using reinforcement learning to align with human preferences.
A practical breakdown of REINFORCE → Actor-Critic → A2C → PPO
If you're working with LLMs, understanding these foundations is essential.
👉 Read the full post: taewoon.kim/2025-08-07-on-po…
💻 Code on GitHub: github.com/tae898/on-policy-…