Sam Hart

dad • engineer • building reasoning infrastructure for AI systems

[Sam Hart @_sam_hart_]

enterprise agents break because they simulate logic instead of enforcing it

[Sam Hart @_sam_hart_]

Session amnesia is a brutal bottleneck on large projects, but this transparency is a feature, not a bug. It shows the gap between stateless models and true persistent memory.

What if we could do cross-episode learning? #LLM #AIAgent #AIEngineering

[Sam Hart @_sam_hart_]

LLMs don't reason. They approximate. Want agents you can trust? Back them with formal logic, ie. datalog, constraint solvers, something with guarantees. Neuro-symbolic isn't a research curiosity. It's the reliability layer your stack is missing.
#NeuroSymbolic #AIAgents

[Sam Hart @_sam_hart_]

Hot take: We keep patching similarity search with rerankers, hybrids, and filters - but the missing piece is logical retrieval.

Think Datalog, not just vectors: intent, context, and task constraints should drive retrieval relevance, not top-k geometry.

[Sam Hart @_sam_hart_]

Thousands of agent memories. Query: "recommend me restaurants in Tokyo". Top-k similarity pulls dining/travel notes. The shellfish allergy? Near-zero rank, low similarity, high relevance. In Japan, shellfish is everywhere. 🦐⚠️

Similarity is not relevance!