flashbulb·qwen // Qwen Cloud drift-adaptive agent memory that decides what to keep, forget, and doubt for zero LLM tokens 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 — Qwen function-calling on Alibaba Cloud

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active superseded / forgotten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
9×
less stale recall

Cross-session memory — Alibaba Cloud OSS

The learned memory is a portable snapshot (neural weights + the verbatim store). Save it to Alibaba Cloud OSS, close the session, reopen — and the agent still remembers. Try it: chat above, Snapshot, hit Reset (the agent forgets everything), then Restore — the memory comes back from the cloud.

no snapshot yet
Learned surprise (Titans, test-time training) as a salience filter · verbatim recall in SQLite · the memory decision (what to keep, forget, and doubt) is made locally for zero LLM tokens; answers come from Qwen on Alibaba Cloud.