We ran LongMemEval-s. 90.79% overall. The two numbers we actually cared about: Knowledge Update 97.43%, Temporal Reasoning 90.97%. Here is why those two, and where we are still soft.
A customer moved last week. Your DB now holds their old address, their new address, two orders shipped to the old one, and a support thread about the move.
Every row is relevant. One address is correct.
Similarity search gets you to the right topic. It cannot tell you which
Swapping the model under an AI product takes an afternoon now. Swapping out everything the product knows about its user takes a rewrite.
Preferences, what the agent already tried and failed at, which of two contradictory facts is the current one, why a decision went the way it
A knowledge graph helps organize what is known.
A context graph helps with decision-making by showing what matters in a specific situation.
That is the simple difference.
A knowledge graph is usually about facts and relationships.
For example:
Customer → has → Contract
