Attention scoring

The attention scorer grades incoming messages 0–100 and assigns a processing level. It's a pure heuristic — no LLM, no network, runs in microseconds.

Use case

In group chats, broadcast channels, or any context where you receive many messages, most don't warrant a full LLM response. The attention scorer tells you whether to engage fully, respond briefly, or skip.

Levels

Level	Score	What to do
full	≥ 75	Engage fully — rich response
standard	≥ 50	Engage — balanced response
minimal	≥ 25	Brief acknowledgement only
ignore	< 25	Skip

Usage

result = mem.score_attention(
    message="URGENT: the API is down!",
    sender="user_123",
    channel="dm",           # "dm" | "group" | "broadcast"
    owner_ids={"user_123"}, # these always get full attention
    allowlist={"user_456"}, # elevated base score
    context={
        "ongoing": True,          # part of an active thread
        "already_answered": False,
    },
)

print(result.score)    # → 85
print(result.level)    # → "full"
print(result.reason)   # → "unknown sender; channel=dm; urgent keyword; question"

Scoring factors

score = sender_score + channel_score + content_score + context_score

Sender: - Owner ID → 100 (immediate full return, no further scoring) - Allowlisted → +60 - Unknown → +10

Channel: - DM → +25 - Group → +15 - Broadcast → +5

Content signals: - Urgent keywords (urgent, asap, emergency, broken, down) → +20 - Action request (do, please, can you, fix, check) → +15 - Question mark → +10 - Casual banter (short message, casual words) → +8 - Single emoji → −5

Context: - Ongoing conversation → +10 - New thread → +10 - Already answered → −15

Configuring thresholds

from extremis import Config, Extremis

mem = Extremis(config=Config(
    attention_full_threshold=80,      # default: 75
    attention_standard_threshold=55,  # default: 50
    attention_minimal_threshold=30,   # default: 25
))

Or via env vars: EXTREMIS_ATTENTION_FULL_THRESHOLD=80

Via MCP

memory_score_attention(
    message="can you help me?",
    sender="user_123",
    channel="group",
    ongoing=true
)

Returns: "Score: 60/100 Level: standard\nReason: ..."