@neo4j@JMHReif Benchmarks are necessary, but scoring final answers misses the best failure signal: which claim was wrong, which source supported it and which definition the system used. RAG gets easier to debug when answers are built from inspectable claims, not opaque chunks.
@Letta_AI Memory benchmarks still overfocus on recall. Production memory also has to answer: is this the same entity, was this true then or now, which source wins, and what conflicts? A model can retrieve the right sentence and still act on the wrong reality.
@kkitase Connecting data without moving it is the easy promise. The real test is whether the catalog can resolve identities across systems, version meaning and surface contradictions. If it cannot, the agent gets unified access to five incompatible definitions of the same thing.
@bcherny Getting agents running is no longer the hard part. The next bottleneck is everything outside the repo: policies, decisions, citizen context, approvals and changing facts. Enterprise deployment needs a shared Knowledge Graph, not 34 separate piles of context.
@sh_reya At scale, the UI cannot just make AI text easier to read. It has to make claims inspectable: which entity is this about, what source supports it, what changed, and what conflicts? Otherwise better formatting only helps humans review generated ambiguity faster.
@cyrilXBT “Search before creating” prevents obvious duplicates but doesn't know that two different notes describe the same person etc. That needs entity resolution. Otherwise Claude becomes a very fast librarian for a random vault.
Try https://t.co/rgkmEAzCwe
@KanikaBK Useful architecture map, but calling it memory stretches the word. A graph of files is not yet a Knowledge Graph. Try https://t.co/rgkmEAzCwe
@mattpavelle Continuous reassessment only works if every conclusion stays tied to its source, timestamp and clinical context. New evidence should not overwrite the old assessment. It should create a new claim and preserve why reality changed. This is a Knowledge Graph problem too.
@sista05 A better schema fixes badly modeled tables. It does not resolve the same entity across five systems, preserve competing claims or explain how a definition changed over time. Schema describes structure. Ontology plus provenance describes meaning and makes it governable.
@0xwhrrari “One employee who knows the company” is the right interface. The hard part is behind it. Slack, CRM and Notion will disagree. Names collide. Decisions change. Entity resolution, provenance, time and conflicts are the layer we’re building with Yggy.
@philhchen Agents are bad at jobs because jobs depend on shared reality, not just execution. Who is the customer? Which decision is current? What did sales promise? What changed? Until agents share an entity-resolved, time-aware model of the company, humans remain the coordination layer.
“RDF or property graph” was the wrong question.
Every fact in our graph is its own node. Metadata sits on the fact: source, confidence, review state, valid_from. The vocabulary on top is RDF/OWL.
Neither camp calls this clean. Wikidata isn’t strictly one or the other either.
@n8n_io@cognee_ Persistent shared memory is necessary, but it is not enough. If every agent can write, the memory quickly fills with duplicates, stale facts and contradictions. Shared identity, provenance, time and conflict handling are what turn storage into reliable memory.
@omarsar0 Episodic memory is great for improving behavior. It is weaker at maintaining a world model. Agents also need to know which entities are identical, which facts changed and which sources disagree. Episodes teach the agent. A Knowledge Graph keeps reality coherent.
@EmanAbio Bounded authority still fails if the agent acts on the wrong version of reality. Before permissions, you need governed state: which entity is this, which claim is current, where did it come from and what conflicts with it. Non-determinism makes semantics operational.
@sermakarevich Great breakdown. One category I’d add: what happens when two memories refer to the same entity under different names, or when both claims were true at different times? Retrieval is the easy part. Entity resolution, temporal truth and conflicts decide whether memory is reliable.
@emollick Multi-agent coordination is getting good fast. The missing bit is shared reality. Passing messages around does not guarantee that every agent means the same thing by a customer, project or decision. Coordination needs a common Knowledge Graph underneath it
@RoundtableSpace Connecting Claude Code and Codex to the same memory is useful. But shared storage is not shared understanding. If they disagree about a customer, project or decision, who resolves it?
@supermemory Claude’s memory is transparent. But md still behaves like md: latest version wins, identity is fuzzy and contradictions disappear into edits. The next step is not better notes. It is a Knowledge Graph that handles ontology preserves claims, sources, time and conflicts.