Research & Findings

The GRAFOMEM Blog

Research findings, technical deep-dives, and the road to verifiable memory for AI agents.

⌫
OPINION RESEARCH
Your agent says it deleted that. It didn't.
I built an AI-memory backend that does something genuinely unsettling: you ask it to delete a fact, it tells you the fact is deleted, its audit log confirms the deletion — and then it hands that exact fact right back to the next query. Every single time.
Camilo Ayerbe · GNS Foundation May 2026 8 min read
⊥
RESEARCH TECHNICAL
A Bigger Embedding Model Won't Fix Your Agent's Memory
'Memory quality' is not one knob — it's four independent levers, and the embedder moves only one. Pull the wrong one and you'll buy a model upgrade that fixes nothing.
GNS Foundation May 2026 6 min read
⬡
TECHNICAL RESEARCH
We Wrote a Formal Conformance Spec for Agent Memory. Most Backends Fail Their Own Claims.
10 capability flags, 135 benchmark traces, 61,754 turns — run against sixteen reference backends. The results: declared ≠ observed.
GNS Foundation May 2026 12 min read
⧖
RESEARCH TECHNICAL
Your Agent Remembers Everything You Told It — Including What's No Longer True
Most agent memory is an append-and-retrieve log that can't express a fact changing. High recall at a generous budget isn't proof it works — it's the flood hiding a missing capability.
GNS Foundation May 2026 6 min read
§
OPINION RESEARCH
AI Memory Needs a Standard — and a Way to Prove You Meet It
The benchmark wars won't settle because the field is measuring the wrong thing. Agent memory doesn't need another leaderboard — it needs a protocol, and a conformance suite that makes 'supports X' mean 'passes the suite for X.'
Camilo Ayerbe May 2026 6 min read
</>
TECHNICAL
Building a GMP-compliant backend in 200 lines
A step-by-step walkthrough of the adapter template — from grafomem init to a passing conformance report.
Coming soon