Retrieval · Reasoning · Execution
Cryptrade does not treat each price tick in isolation. A Retrieval-Augmented Generation (RAG) knowledge base grows automatically from live feeds, Hermes AI digests, and trade outcomes — then the best-matching facts are injected into every strategy and gate evaluation before a swap is signed.
Knowledge flows in continuously; decisions pull the most relevant facts out at runtime.
memoryBrief)Chronological log of this session: ticks, signals, skips, fills, AI verdicts, news hits.
data/bot-memory.jsonlknowledgeBrief)Durable, tagged facts with confidence scores and TTL — the RAG corpus.
data/market-knowledge.jsonlAt every AI evaluation, Hermes receives both briefs plus live quote, EMA, order book, macro, and intel context. The strategy math fires first; RAG shapes whether gates pass and how confident the AI is.
Every ~5 minutes collectors upsert fresh chunks. Hourly digest adds Hermes-distilled facts. Outcomes close the loop.
| Source | What gets written | TTL (typical) | Role in trades |
|---|---|---|---|
seed |
Baseline rules: wallet hygiene, accumulation scenarios, exchange references | Permanent | Always-retrieved guardrails |
intel_feed |
Fear & Greed, funding rates, spot references, composite intel score | 12–72 h | Regime context; adjusts AI confidence discount |
macro_feed |
Crypto vs metals leadership, risk-on / risk-off regime | 48 h | Macro gate + bearish accumulation caution |
orderbook_feed |
Binance L2 imbalance, spread, mid prices | 6 h | Order book gate; CEX vs DEX sanity |
wallet |
Sanitized live wallet: USDC balance, gas, active chain | 24 h | Size and readiness checks |
whale_feed |
Large transfer bias summary | 24 h | Whale gate on BUY/SELL |
news_feed |
Verified headlines (BBC, CNBC, WSJ, Bloomberg, FT, Fed, SEC) | 48 h | Event playbooks; desk intel |
ai_learning |
Win rate, blocked signals, correlated AI vs outcome patterns | 168 h | Teaches Hermes what blocks saved capital |
hermes_digest |
1–3 new facts distilled hourly from session + feeds | 72 h | Durable learned cautions and scenarios |
hermes_gate |
Snapshot of each AI assessment at gate time | 48 h | Audit trail; future RAG retrieval |
When a BUY or SELL signal arrives, retrieveKnowledgeChunks scores every active chunk and builds knowledgeBrief.
+25 for chain tag (e.g. ethereum). +18 per matching asset (eth, weth, usdc). Focus accumulation boosts scenario and pattern categories.
Market regime and pattern chunks written in the last 6 hours get up to +22. Stale intel naturally fades as TTL expires.
Security and core accumulation scenarios are always included in the brief — even if they would not top the score alone.
Each chunk carries a 0–100 confidence score. Seed facts and wallet snapshots rank high; Hermes digest facts start at ~70 until reinforced by outcomes.
Price websocket triggers EMA dip logic, take-profit, or stop-loss. No AI required for the initial signal — math defines when to consider a trade.
knowledgeBriefForContext retrieves top chunks for chain, pair (WETH/USDC), and accumulation focus. Session memoryBrief adds recent skips, fills, and AI history.
AiAdvisor.getIntelligence sends briefs + live context to Ollama. Returns bias, confidence, action_hint, risk flags, and invalidation — logged to session memory.
When AI_GATE_BUY / AI_GATE_SELL are on, Hermes must approve. CEX, macro, order book, and whale gates also consult live feeds already mirrored in RAG. Failed gate → skip with desk commentary.
Executed trades and skipped signals get correlated by signal ID. Win rates and “AI blocked N signals” patterns upsert into ai_learning chunks — retrieved on the next decision.
AI_MIN_CONFIDENCEWith AI gates off, RAG still enriches desk and hourly digest — strategy may execute if other gates pass.
refreshKnowledgeBase runs collectors: wallet, intel, macro, orderbook, whale, news, memory learning. Fresh snapshots replace expired TTL chunks.
digestMarketKnowledge asks Hermes for 1–3 new JSON facts given current memory + RAG. Upserts as hermes_digest with 72 h TTL.
Ranked chunks compress into knowledgeBrief (~1800 chars). Injected into getIntelligence alongside macro and intel briefs.
SignalOutcomeTracker links AI verdicts to PnL. collectFromMemoryLearning writes pattern chunks so future trades benefit from closed-loop feedback.
| Env variable | Default | Effect |
|---|---|---|
KNOWLEDGE_ENABLED | true | Master switch for RAG store |
KNOWLEDGE_COLLECT_SEC | 300 | How often collectors refresh feeds |
KNOWLEDGE_DIGEST_SEC | 3600 | Hermes hourly fact distillation |
KNOWLEDGE_BRIEF_MAX_CHARS | 1800 | Max RAG text injected per decision |
AI_GATE_BUY / AI_GATE_SELL | configurable | Require Hermes approval before swap |
AI_MIN_CONFIDENCE | — | Minimum % for BUY approval (adjusted by intel) |
Monitor live facts locally: npm run rag:monitor:open
Session tape vs knowledge store, privacy, and what AI does not control.
Every external feed that feeds the collectors above.
Execution, safety checks, and on-chain audit trail.
See signals, intel, and gate blocks in real time.