williamcorrea23 SAP Self-Learn — Hermes-Style Environment Adaptation
Builds internal knowledge graph from every SAP interaction. Like Hermes agent but focused on SAP development workflows — MCP performance, routing success, system quirks, user preferences.
Philosophy (Andrej Karpathy)
measure → learn → adapt → repeat
Never route blindly. Every decision backed by data. Track what works, deprecate what fails, discover what's possible.
Architecture
Every MCP Call
│
▼ Timestamp, latency, success/fail, error
┌──────────────────────────────┐
│ Self-Learn Engine │
│ │
│ 1. Record observation │
│ 2. Update EMA statistics │
│ 3. Adapt routing weights │
│ 4. Persist to MEMORY.md │
│ ## LEARN section │
└──────────┬───────────────────┘
│
▼ Context injected into every prompt
┌──────────────────────────────────────────────┐
│ MCP_RELIABILITY: arc-1=98% aibap=95% ... │
│ SYS: hana=2.0 s4h=2023 gcts=true ... │
│ PATTERNS: material_create→ZROUTER ... │
└──────────────────────────────────────────────┘
Integration with sap_router.py
# After every routing decision:
from scripts.self_learn import SelfLearnEngine
engine = SelfLearnEngine()
engine.load_history()
# Record MCP call outcome
engine.record_mcp_call(
mcp_name='arc-1',
latency_ms=245,
success=True
)
# Adapt route based on history
adapted = engine.adapt_route('MM_CREATE_MATERIAL', primary_route)
if adapted.get('confidence', 0.5) < 0.3:
# Low confidence — try alternative
...
# Persist learned context
engine.persist()
CLI Commands
# Record MCP call outcome
python scripts/self_learn.py record-mcp --mcp arc-1 --latency 245 --success true
# Record routing outcome
python scripts/self_learn.py record-route --action MM_CREATE_MATERIAL --success true
# Get best MCP among candidates
python scripts/self_learn.py best-mcp --candidates "arc-1,aibap,mcp-abap-adt"
# Get learned context for prompt injection
python scripts/self_learn.py context
# Force persist to MEMORY.md
python scripts/self_learn.py persist
# Discover SAP system feature
python scripts/self_learn.py discover --feature hana_version --value "2.00.080.00"
MEMORY.md LEARN Section Format
## LEARN
- sys:adt_version 2.115
- sys:basis_release 757
- sys:gcts true
- sys:hana_version 2.00.080.00
- sys:rap_version managed
- mcp:aibap latency:180ms success:0.92 last:2026-06-26T14:30:00
- mcp:arc-1 latency:245ms success:0.98 last:2026-06-26T14:35:00
- mcp:mcp-abap-adt latency:350ms success:0.85 last:2026-06-26T14:20:00
- mcp:mcp-sap-gui latency:1200ms success:0.75 last:2026-06-26T12:00:00
- route:MM_CREATE_MATERIAL total:45 ok:43 fail:2
- route:BASIS_CODE_SEARCH total:30 ok:28 fail:2
- pattern:mass_material_create ZROUTER prefer_batch_for_50plus
- pattern:quick_source_read ADT prefer_direct_for_single_classes
Auto-Discovery
Self-learn engine automatically discovers:
| Feature | Detection Method |
|---|---|
| ADT version | Parse arc-1 SAPDiagnose output |
| Basis release | Parse aibap system_info output |
| HANA version | Parse sap_hana_query SELECT * FROM M_DATABASE |
| gCTS available | Probe aibap gcts_list success |
| RAP version | Probe CDS view with managed scenario |
| BTP subaccount | Parse btp-mcp list_subaccounts |
| CPI tenant | Probe sap-cpi list_packages |
Routing Adaptation
When confidence drops below threshold, self-learn suggests alternatives:
Primary route confidence: 0.25 (5 failures in 20 attempts)
Alternatives:
1. ADT direct (confidence: 0.92) — recommended
2. GUI fallback (confidence: 0.85)
3. ZROUTER RFC (confidence: 0.45) — same failure pattern
Gotchas
- Cold start: No history → neutral routing (no bias). Confidence stays 0.5 until 3+ observations.
- Decay: Stats decay over 24h half-life. Recent performance matters more.
- No persistence on error: Failed calls recorded but don't corrupt stats. 30% EMA weighting.
- Context injection: Learned context auto-injected into every LLM prompt for routing decisions. ~50 tokens overhead.
- Memory budget: LEARN section capped at 50 lines. Older stats archived as ARCHIVE_LEARN.