Graph & Training

Training Signals & Feedback Loops

What counts as high-signal memory differs between a Rust systems team, a Rails fintech team, and a legal research group. Every day-to-day curation action in Sulcus generates tenant-isolated training signals that specialize SIU and System 1 to your domain.

SIVU + SICU SignalsTenant-Isolated ModelsSIRU Recall AnalyticsZero Cross-Tenant Leakage
01 · Signal Collection

How Everyday Actions Generate Training Data

You don't need to build separate labeling spreadsheets. Agents and humans emit structured signals naturally while working.
Action / SourceTarget UnitEffect on Future Curation
memory_reclassify (MCP / Studio)SICU Archetype HeadLogs a high-confidence ground-truth archetype correction whenever an agent or human changes memory_type.
POST /api/v1/feedbackSIVU + SIRU"relevant" boosts salience and reinforces store utility; "irrelevant" or "outdated" dampens heat and logs negative quality feedback.
memory_deprecate / DeleteSIVU Quality HeadDeprecating or deleting noisy auto-captured turns teaches SIVU to reject similar low-value patterns in your domain.
Pin / High-Frequency RecallSIVU + StabilityPinned nodes and repeatedly paged-in territories serve as strong positive quality examples during retraining.
POST /api/v1/siu/signalSIVU + SICUDirect API and Studio labeling of raw text samples with quality (store | reject) and corrected_type.
POST /api/v1/triggers/feedbackSITU PrecisionRecords helpful, noisy, dismissed, acted_on, or false_positive signals on fired triggers.
02 · Programmatic Feedback

Recording Signals & Triggering Retrains

Custom System 1 weights resolve in strict hierarchy: (tenant_id, namespace) -> (tenant_id, __global__) -> compiled-in baseline weights.
pythonPython SDK Feedback & Signals
from sulcus import Sulcus

client = Sulcus(api_key="sk-...")

# 1. Explicit recall relevance feedback (updates heat + records SIRU/SIVU signal)
client.feedback("0195f410-8a11-7000-8000-000000000010", "relevant")

# 2. Record a direct System 1 labeling signal (store/reject + archetype)
client.record_siu_signal(
    text="Run bundle exec rspec spec/requests/ before cutting a release tag",
    quality_label="store",
    corrected_type="procedural",
    namespace="payments-rails",
)
typescriptTypeScript & Node.js SDK (sulcus)
import { Sulcus } from "sulcus";

const client = new Sulcus({ apiKey: process.env.SULCUS_API_KEY! });

// Reclassify a memory node over REST/SDK and automatically log a SICU training signal
await client.reclassifyMemory({
  nodeId: "0195f410-8a11-7000-8000-000000000010",
  newType: "fact",
  reason: "Verified production database schema invariant",
});
bashInspecting Status & Running In-Process Pure-Rust Retraining
# Inspect accumulated training signals and active System 1 model metrics
curl https://api.sulcus.ca/api/v1/siu/status \
  -H "Authorization: Bearer $SULCUS_API_KEY"

# Trigger tenant-isolated pure-Rust System 1 retraining and atomic hot-reload
curl -X POST https://api.sulcus.ca/api/v1/siu/retrain \
  -H "Authorization: Bearer $SULCUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"namespace": "payments-rails"}'
Held-Out Validation Gate & Anchor Replay
Every POST /api/v1/siu/retrain run blends your accumulated tenant signals with a compiled-in anchor replay pack so fine-tuning on a small domain sample never causes catastrophic forgetting. New weights hot-reload into memory only if they pass the held-out validation gate.
03 · Protocol Reference

Training & Analytics Endpoints

POST/api/v1/siu/signal
Record an explicit SIVU quality or SICU archetype training signal.
GET/api/v1/siu/signals
List stored training signals (?namespace, ?limit).
GET/api/v1/siu/status
Inspect active System 1 model metrics and pending signal counts.
POST/api/v1/siu/retrain
Run pure-Rust System 1 training and hot-reload promoted weights.
GET/api/v1/siu/models
List custom tenant and namespace System 1 models.
PUT/api/v1/siu/models
Upload offline-trained System 1 weights.
DELETE/api/v1/siu/models
Delete custom weights and revert to baseline System 1 weights.
POST/api/v1/feedback
Record recall relevance feedback (relevant | irrelevant | outdated).
GET/api/v1/analytics/recall
Inspect tenant recall hit rates and similarity distributions.
CLI & Offline Training (`sulcus train`)