File size: 10,807 Bytes
b96103d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 | """
Sentinel Memory Engine — Temporal context storage and trend analysis.
Stores periodic snapshots of the ContextState, creating a timeline of what
the system has observed. This enables:
- Trend detection: "situation is getting worse over the last 30 seconds"
- Historical queries: "what was happening 2 minutes ago?"
- VLM context enrichment: "previous alerts show a person approaching"
- Escalation logic: "3 warnings in 30 seconds → auto-escalate to critical"
Usage:
memory = MemoryEngine(max_entries=100)
memory.store(context_engine.state, alert_level="warning", alert_text="Person nearby")
summary = memory.get_trend_summary() # for VLM prompt
trend = memory.analyze_trend() # "deteriorating" / "improving" / "stable"
"""
import time
from dataclasses import dataclass, field
try:
import structlog
logger = structlog.get_logger()
except ImportError:
import logging
logger = logging.getLogger("sentinel.memory")
from context_engine import ContextState
@dataclass
class MemoryEntry:
"""
A point-in-time snapshot of the system's understanding.
Combines the context state with any alert that was generated,
creating a complete record of "what happened at this moment."
"""
timestamp: float
context_summary: str
activity: str
environment: str
location_type: str
nearby_people: int
closest_person_distance: float
closest_person_trend: str
risk_level: float
risk_factors: list = field(default_factory=list)
alert_level: str = "ok"
alert_text: str = ""
battery: float = 100.0
frames_analyzed: int = 0
@classmethod
def from_context(
cls,
ctx: ContextState,
alert_level: str = "ok",
alert_text: str = ""
) -> "MemoryEntry":
"""Create a MemoryEntry from a live ContextState snapshot."""
return cls(
timestamp=time.time(),
context_summary=ctx.to_prompt_context(),
activity=ctx.activity,
environment=ctx.environment,
location_type=ctx.location_type,
nearby_people=ctx.nearby_people,
closest_person_distance=ctx.closest_person_distance,
closest_person_trend=ctx.closest_person_trend,
risk_level=ctx.risk_level,
risk_factors=list(ctx.risk_factors),
alert_level=alert_level,
alert_text=alert_text,
battery=ctx.battery,
frames_analyzed=ctx.frames_analyzed,
)
def age_seconds(self) -> float:
return time.time() - self.timestamp
def is_alert(self) -> bool:
return self.alert_level in ("warning", "critical")
class MemoryEngine:
"""
Bounded timeline of context snapshots for temporal reasoning.
Stores up to max_entries (default 100) MemoryEntry objects.
Provides trend analysis, historical queries, and formatted output
for VLM prompt enrichment.
Designed for Gradio per-session state (gr.State). All attributes
are plain Python types — fully serializable.
"""
def __init__(self, max_entries: int = 100):
self.entries: list[MemoryEntry] = []
self.max_entries = max_entries
self._last_store_time: float = 0.0
def store(
self,
context: ContextState,
alert_level: str = "ok",
alert_text: str = "",
min_interval: float = 2.0
) -> MemoryEntry | None:
"""
Store a context snapshot.
Args:
context: Current ContextState to snapshot.
alert_level: Alert level if one was generated this frame.
alert_text: Alert text if one was generated.
min_interval: Minimum seconds between stores (prevents flooding).
Alerts (warning/critical) bypass this throttle.
Returns:
The created MemoryEntry, or None if throttled.
"""
now = time.time()
is_alert = alert_level in ("warning", "critical")
if not is_alert and (now - self._last_store_time) < min_interval:
return None
entry = MemoryEntry.from_context(context, alert_level, alert_text)
self.entries.append(entry)
if len(self.entries) > self.max_entries:
self.entries = self.entries[-self.max_entries:]
self._last_store_time = now
return entry
def recent(self, seconds: float = 60.0) -> list[MemoryEntry]:
"""Retrieve entries from the last N seconds."""
cutoff = time.time() - seconds
return [e for e in self.entries if e.timestamp >= cutoff]
def recent_alerts(self, seconds: float = 120.0) -> list[MemoryEntry]:
"""Retrieve only alert entries (warning/critical) from last N seconds."""
cutoff = time.time() - seconds
return [e for e in self.entries if e.timestamp >= cutoff and e.is_alert()]
def alert_count(self, seconds: float = 60.0, level: str | None = None) -> int:
"""Count alerts in the last N seconds, optionally filtered by level."""
cutoff = time.time() - seconds
count = 0
for e in self.entries:
if e.timestamp < cutoff:
continue
if not e.is_alert():
continue
if level is None or e.alert_level == level:
count += 1
return count
# ── Trend Analysis ──────────────────────────────────────────────────────
def analyze_trend(self, window: float = 30.0) -> dict:
"""
Compare recent entries against older entries to detect trends.
Returns a dict:
{
"direction": "deteriorating" | "improving" | "stable",
"risk_delta": float, # positive = risk increasing
"alert_rate": float, # alerts per minute
"warning_count": int, # warnings in window
"critical_count": int, # criticals in window
"person_approaching": bool, # is someone closing in?
"activity_change": str | None, # e.g. "walking → fallen"
}
"""
now = time.time()
recent = [e for e in self.entries if e.timestamp > now - window]
older = [
e for e in self.entries
if now - window * 2 < e.timestamp <= now - window
]
result = {
"direction": "stable",
"risk_delta": 0.0,
"alert_rate": 0.0,
"warning_count": 0,
"critical_count": 0,
"person_approaching": False,
"activity_change": None,
}
if len(recent) < 2:
return result
recent_risk = sum(e.risk_level for e in recent) / len(recent)
older_risk = (
sum(e.risk_level for e in older) / len(older) if older else 0.0
)
risk_delta = recent_risk - older_risk
result["risk_delta"] = risk_delta
if risk_delta > 0.15:
result["direction"] = "deteriorating"
elif risk_delta < -0.15:
result["direction"] = "improving"
alerts_in_window = [e for e in recent if e.is_alert()]
result["warning_count"] = sum(
1 for e in alerts_in_window if e.alert_level == "warning"
)
result["critical_count"] = sum(
1 for e in alerts_in_window if e.alert_level == "critical"
)
result["alert_rate"] = len(alerts_in_window) / (window / 60.0)
if any(e.closest_person_trend == "approaching" for e in recent):
result["person_approaching"] = True
if older and recent:
old_activity = max(set(e.activity for e in older), key=lambda a: sum(1 for e in older if e.activity == a))
new_activity = max(set(e.activity for e in recent), key=lambda a: sum(1 for e in recent if e.activity == a))
if old_activity != new_activity:
result["activity_change"] = f"{old_activity} → {new_activity}"
return result
# ── Formatted Output ────────────────────────────────────────────────────
def get_trend_summary(self, max_entries: int = 5) -> str:
"""
Format recent history for VLM prompt enrichment.
Returns a compact multi-line string with timestamps and alerts.
"""
recent = self.entries[-max_entries:]
if not recent:
return "No previous context available."
lines = ["Recent observations:"]
for entry in recent:
t = time.strftime("%H:%M:%S", time.localtime(entry.timestamp))
alert_tag = f" [{entry.alert_level.upper()}]" if entry.is_alert() else ""
alert_msg = f" — {entry.alert_text}" if entry.alert_text else ""
lines.append(
f" [{t}]{alert_tag} {entry.activity} in {entry.environment} "
f"{entry.location_type}, risk={entry.risk_level:.2f}{alert_msg}"
)
return "\n".join(lines)
def get_situation_summary(self, seconds: float = 120.0) -> str:
"""
High-level narrative of what happened in the last N seconds.
Suitable for incident reports or caregiver dashboards.
"""
entries = self.recent(seconds)
if not entries:
return "No activity recorded."
alerts = [e for e in entries if e.is_alert()]
avg_risk = sum(e.risk_level for e in entries) / len(entries)
activities = list(set(e.activity for e in entries))
trend = self.analyze_trend(seconds)
parts = [
f"Over the last {seconds:.0f}s:",
f" {len(entries)} observations, {len(alerts)} alerts",
f" Average risk: {avg_risk:.2f}",
f" Activities: {', '.join(activities)}",
f" Trend: {trend['direction']}",
]
if trend["activity_change"]:
parts.append(f" Activity changed: {trend['activity_change']}")
if trend["person_approaching"]:
parts.append(" Person detected approaching")
if trend["critical_count"] > 0:
parts.append(f" {trend['critical_count']} critical alerts fired")
return "\n".join(parts)
def reset(self) -> None:
"""Clear all stored entries."""
self.entries.clear()
self._last_store_time = 0.0
@property
def total_entries(self) -> int:
return len(self.entries)
@property
def total_alerts(self) -> int:
return sum(1 for e in self.entries if e.is_alert())
|