flashbulb·amd // MI300X open, self-hostable agent memory — decides what to keep for zero LLM tokens, nothing leaves the box linking…
Surprise signal each turn fires a surprise; above the trigger τ it burns in as memory (0 tokens), below it decays unheard.
Write-path tokens / turn — cost of deciding what to remember
0vs—
llm_extract baseline (what Mem0/Zep/Letta spend) vs our surprise gate
Eliminated
—
run a preset →
Decision latency
~5ms
~37× faster than one LLM call

Live agent — Gemma 4 · on the MI300X

Say something, or run a preset →

Retention — a knowledge update forgets the stale fact

active superseded / forgotten
0
tokens / write
2.8M
params · learning
96.7%
gate @ τ=0.40
1
MI300X · full stack
Learned surprise (Titans, test-time training) as a salience filter · verbatim recall in SQLite · memory compute and the Gemma 4 LLM co-located on one AMD Instinct MI300X via ROCm — zero external inference, nothing leaves the box.