envelope in / out — LLM becomes a callable function
4
Provider-swappable
hermes model swaps GLM · Opus · GPT in one line
5
Closed learning loop
MEMORY.md auto-injected before the envelope — part of Hermes' self-improving runtime
§ 04Live Demo
Watch MEMORY.md rewrite itself.
demo · MEMORY.md closed loop · live capture
Run 1 writes an operational prior into MEMORY.md.
Run 2 picks it up before our envelope arrives — behavior shifts,
but TraceGuard still gates the answer.
Fig. — closed loop
Run 1finishes — Hermes writes operational prior to MEMORY.md
writes
MEMORY.mdpersistent across runs — agent-curated, on_memory_write hook
auto-prepended to system prompt
Run 2starts with prior already in system prompt, before our envelope
answer flows down
TraceGuardstill gates — evidence must come from this run's fresh child manifest
§ 05How TraceGuard Caught Our Paper
We fed our own paper as input — and caught a real bug.
1
input
Paper sections
paper key sections target
2
decompose
Ouroboros
→ 4 bounded chunks
c0
c1
c2
c3
3
Hermes children
4 child calls
→ child_result : 000..003
000
001
002
003
4
parent synthesis
cites child : 004
— does not exist
000001002003004
5 · TraceGuardREJECT
Raw parent rejectedunsupported_fact_id · child_result : 004
0.0435
6 · Repair & retryACCEPT
Handle ↦ fresh manifestcites only child_result : 000..003