Agent Brain(memory)Research

Context the whole company can use.

Can one brain improve the whole workforce?

Remembering is not retrieving

Most “agent memory” is retrieval: text in, similar text out. Useful, but not the same problem — an agent that worked with you last month should know what was decided, what changed, and what is no longer true. We model memory as claims with two clocks (when it was true, when we believed it), so a store can answer what it believed last Tuesday and a correction can supersede a fact without erasing that it was once held.

A negative result we keep repeating

The intuitive design — a model reading every conversation and writing structured facts — did not earn its cost, and the failure is structural: extraction decides what to keep before anyone knows what will be asked, and what it drops is unrecoverable. The lever that worked sits on the read side: how you interrogate the store at the moment of the question beat anything we did at write time.

Benchmarks

Recall no other memory system reaches.

FictionalCharacters QA on a matched gpt-4o-mini backbone. 22 published systems answer ≤7% of multi-hop questions correctly — Brain answers 57%.

SystemSingle-hopMulti-hopBackbone
Naïve Brain91%57%gpt-4o-mini
CAR (prior SOTA)78%30.2%gpt-4o-mini
HippoRAG-v254%≤7%
MemGPT / Cognee28%≤7%
Mem0 / Contriever18%≤7%
Zep / Graphiti7%≤7%

FictionalCharacters QA · % of held-out questions answered correctly, higher is better · Naïve measured, others published · July 2026

Paper
  • Coming soonBoth Engines, One Harness: A Self-Run Head-to-Head Against mem0's Open-Source Server
About this note
Status
Research
Focus
Cost and usefulness (research)
Surface
Early / evolving
Papers
Both Engines, One Harness (preprint)

Follow the work

Early Lab. Papers and findings land here when they clear the gate.