Narrative
Why I test agent memory
The starting point for this lab: Hermes, MCP, Codex, Claude Code, then the conviction that memory is becoming a product layer.
I started with Hermes
Hermes got me hooked because it points at a concrete problem. A model sets the ceiling, but an agent really lives or dies on what it holds onto, what it drops, and whatever it can dig back up once a task runs long.
The more time I spend with these tools, the more I care about memory specifically. Not the vague marketing kind that promises to remember everything you ever typed. I mean the operational stuff: decisions and preferences, entities and their history, corrections, plus the unglamorous parts like isolation, security, and getting your data back after a crash.
Sibyl became my test bench
Sibyl is a useful case study because the repo contains real suites: memory benchmarks, competitor comparisons, MCP security tests, Hermes tests, audit reports and rerun scripts.
I am not selling these results as universal truth, just as test work. Every number links back to a file, the limits stay in plain sight, and each comparison spells out why some plugins were never run in their most expensive configuration.
The important point
Who has the best score is the boring question. What I actually want to understand is the real cost of memory: how long ingestion takes, how many tokens get shipped to the model, how big the average context grows, whether profiles stay isolated, and how wide the MCP attack surface gets.
That is what this site has to make readable, from the story down to the tests, the evidence, and the scripts you can run yourself.