Configuration

All settings live in MemgarConfig and can be set via YAML, environment variables, or constructor args.

Environment variables

Variable	Default	Effect
MEMGAR_FEED_ENABLED	true	Pull signed threat feed at startup
MEMGAR_FEED_MAX_AGE_DAYS	7	Stale cache age before re-fetch
MEMGAR_FEED_GITHUB_REPO	slcxtor/memgar	Where to fetch the signed bundle
MEMGAR_OBSERVABILITY_ENABLED	false	Run Prometheus + drift monitor
MEMGAR_OBSERVABILITY_PORT	9090	Prometheus scrape port
MEMGAR_OBSERVABILITY_DRIFT_THRESHOLD	0.20	PSI threshold for drift alerts
MEMGAR_OBSERVABILITY_DRIFT_WINDOW	1000	Window of scans for drift baseline
MEMGAR_CACHE_DIR	~/.cache/memgar	Pattern cache + feed cache root
MEMGAR_FAIL_CLOSE	false	Escalate ALLOW → QUARANTINE when ML/feed degraded
MEMGAR_TRANSFORMER_THRESHOLD	0.92	Min ONNX prob to fire ML-DETECT signal

YAML config

# memgar.yaml
analysis:
  fail_close: true                 # MEMGAR_FAIL_CLOSE equivalent
  use_llm: false
  semantic_guard: true
  use_transformer_ml: true

feed:
  enabled: true
  max_age_days: 7
  github_repo: slcxtor/memgar

observability:
  enabled: true
  port: 9090
  drift_threshold: 0.2
  drift_window: 1000

Load with:

from memgar.config import MemgarConfig
cfg = MemgarConfig.from_yaml("memgar.yaml")

Constructor overrides

Constructor args take precedence over both YAML and env vars:

from memgar import Analyzer

a = Analyzer(
    use_llm=False,
    fail_close=True,             # overrides MEMGAR_FAIL_CLOSE
    semantic_guard=True,
    use_transformer_ml=True,
    strict_mode=False,
    use_whitelist=True,
    use_sliding_window=True,
    window_size=1000,
    window_overlap=200,
    correlation_detector=True,
    ensemble_voter=True,
    similarity_layer=True,
)

fail_close decision

flowchart LR
    A[Analyze] --> B{decision}
    B -->|ALLOW| C{fail_close & any layer degraded?}
    C -->|yes| D[Decision.QUARANTINE]
    C -->|no| E[Decision.ALLOW]
    B -->|QUARANTINE / BLOCK| F[unchanged]

When fail_close fires, the explanation is prefixed with [fail_close] Escalated ALLOW→QUARANTINE: 2 layer(s) degraded (layer1.5_semantic_guard, threat_feed). ... so operators can see exactly which signal was missing.

Per-source trust

a.register_source_trust("openai-api",      0.95)
a.register_source_trust("user-form",       0.40)
a.register_source_trust("untrusted-wiki",  0.10)

Trust is purely operator-defined — memgar never silently learns trust from behaviour. Auto-learned trust would be its own attack surface.