Skip to content
memseek
Writing

Notes on agent memory.

Why similarity retrieval serves stale answers, how to date and trace context, and what we learn building Memseek.

Agent architecture 2Agent memory 2Context engineering 2Memory 1Open source 1Retrieval 1Verifiability 1World models 1
A memory stream, a four-level hierarchy, a knowledge graph, and a reflection loop connect to the same eight-module context-engine backbone.

Everyone Is Building the Same Context Engine

Most agent products quietly rebuild the same backend: collect evidence, maintain state, retrieve what matters, assemble context, and learn from outcomes. We open-sourced that layer as Memseek.

Context engineeringAgent architectureOpen source Read →

The Agent Loop Has Always Been There

Why verifiability, world models, and memory matter—and why agents need a context engine around the model

Agent architectureContext engineeringVerifiability Read →
A timeline of three decisions, two marked superseded, with only the third served to the model.

Your agent doesn't need more context. It needs the current context.

Similarity retrieval has no opinion about what's still true. That's not a ranking problem you can tune away — it's a missing edge in your data.

Agent memoryRetrieval Read →