mirror of
https://github.com/openclaw/openclaw.git
synced 2026-05-28 06:50:42 +00:00
08b08c66f1
Open proposals ("Want me to...?", "Should I...?") are dangerous in
long-term memory because other sessions interpret them as active
instructions and attempt to carry them out. This adds:
- Attention gate patterns for cron delivery outputs and assistant proposals
- Extractor scoring rules to rate proposals/action items as low importance
- Sleep-cycle Phase 7 to retroactively clean existing noise-pattern memories
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
253 lines
10 KiB
TypeScript
253 lines
10 KiB
TypeScript
/**
|
|
* Attention gate — lightweight heuristic filter (phase 1 of memory pipeline).
|
|
*
|
|
* Rejects obvious noise without any LLM call, analogous to how the brain's
|
|
* sensory gating filters out irrelevant stimuli before they enter working
|
|
* memory. Everything that passes gets stored; the sleep cycle decides what
|
|
* matters.
|
|
*/
|
|
|
|
const NOISE_PATTERNS = [
|
|
// Greetings / acknowledgments (exact match, with optional punctuation)
|
|
/^(hi|hey|hello|yo|sup|ok|okay|sure|thanks|thank you|thx|ty|yep|yup|nope|no|yes|yeah|cool|nice|great|got it|sounds good|perfect|alright|fine|noted|ack|kk|k)\s*[.!?]*$/i,
|
|
// Two-word affirmations: "ok great", "sounds good", "yes please", etc.
|
|
/^(ok|okay|yes|yeah|yep|sure|no|nope|alright|right|fine|cool|nice|great)\s+(great|good|sure|thanks|please|ok|fine|cool|yeah|perfect|noted|absolutely|definitely|exactly)\s*[.!?]*$/i,
|
|
// Deictic: messages that are only pronouns/articles/common verbs — no standalone meaning
|
|
// e.g. "I need those", "let me do it", "ok let me test it out", "I got it"
|
|
/^(ok[,.]?\s+)?(i('ll|'m|'d|'ve)?\s+)?(just\s+)?(need|want|got|have|let|let's|let me|give me|send|do|did|try|check|see|look at|test|take|get|go|use)\s+(it|that|this|those|these|them|some|one|the|a|an|me|him|her|us)\s*(out|up|now|then|too|again|later|first|here|there|please)?\s*[.!?]*$/i,
|
|
// Short acknowledgments with trailing context: "ok, ..." / "yes, ..." when total is brief
|
|
/^(ok|okay|yes|yeah|yep|sure|no|nope|right|alright|fine|cool|nice|great|perfect)[,.]?\s+.{0,20}$/i,
|
|
// Conversational filler / noise phrases (standalone, with optional punctuation)
|
|
/^(hmm+|huh|haha|ha|lol|lmao|rofl|nah|meh|idk|brb|ttyl|omg|wow|whoa|welp|oops|ooh|aah|ugh|bleh|pfft|smh|ikr|tbh|imo|fwiw|np|nvm|nm|wut|wat|wha|heh|tsk|sigh|yay|woo+|boo|dang|darn|geez|gosh|sheesh|oof)\s*[.!?]*$/i,
|
|
// Single-word or near-empty
|
|
/^\S{0,3}$/,
|
|
// Pure emoji
|
|
/^[\p{Emoji}\s]+$/u,
|
|
// System/XML markup
|
|
/^<[a-z-]+>[\s\S]*<\/[a-z-]+>$/i,
|
|
|
|
// --- Session reset prompts (from /new and /reset commands) ---
|
|
/^A new session was started via/i,
|
|
|
|
// --- Raw chat messages with channel metadata (autocaptured noise) ---
|
|
/\[slack message id:/i,
|
|
/\[message_id:/i,
|
|
/\[telegram message id:/i,
|
|
|
|
// --- System infrastructure messages (never user-generated) ---
|
|
// Heartbeat prompts
|
|
/Read HEARTBEAT\.md if it exists/i,
|
|
// Pre-compaction flush prompts
|
|
/^Pre-compaction memory flush/i,
|
|
// System timestamp messages (cron outputs, reminders, exec reports)
|
|
/^System:\s*\[/i,
|
|
// Cron job wrappers
|
|
/^\[cron:[0-9a-f-]+/i,
|
|
// Gateway restart JSON payloads
|
|
/^GatewayRestart:\s*\{/i,
|
|
// Background task completion reports
|
|
/^\[\w{3}\s+\d{4}-\d{2}-\d{2}\s.*\]\s*A background task/i,
|
|
|
|
// --- Conversation metadata that survived stripping ---
|
|
/^Conversation info\s*\(/i,
|
|
/^\[Queued messages/i,
|
|
|
|
// --- Cron delivery outputs & scheduled reminders ---
|
|
// Scheduled reminder injection text (appears mid-message)
|
|
/A scheduled reminder has been triggered/i,
|
|
// Cron delivery instruction to agent (summarize for user)
|
|
/Summarize this naturally for the user/i,
|
|
// Relay instruction from cron announcements
|
|
/Please relay this reminder to the user/i,
|
|
// Subagent completion announcements (date-stamped)
|
|
/^\[.*\d{4}-\d{2}-\d{2}.*\]\s*A sub-?agent task/i,
|
|
// Formatted urgency/priority reports (email summaries, briefings)
|
|
/(\*\*)?🔴\s*(URGENT|Priority)/i,
|
|
// Subagent findings header
|
|
/^Findings:\s*$/im,
|
|
// "Stats:" lines from subagent completions
|
|
/^Stats:\s*runtime\s/im,
|
|
];
|
|
|
|
/** Maximum message length — code dumps, logs, etc. are not memories. */
|
|
const MAX_CAPTURE_CHARS = 2000;
|
|
|
|
/** Minimum message length — too short to be meaningful. */
|
|
const MIN_CAPTURE_CHARS = 30;
|
|
|
|
/** Minimum word count — short contextual phrases lack standalone meaning. */
|
|
const MIN_WORD_COUNT = 8;
|
|
|
|
/** Shared checks applied by both user and assistant attention gates. */
|
|
function failsSharedGateChecks(trimmed: string): boolean {
|
|
// Injected context from the memory system itself
|
|
if (trimmed.includes("<relevant-memories>") || trimmed.includes("<core-memory-refresh>")) {
|
|
return true;
|
|
}
|
|
|
|
// Noise patterns
|
|
if (NOISE_PATTERNS.some((r) => r.test(trimmed))) {
|
|
return true;
|
|
}
|
|
|
|
// Excessive emoji (likely reaction, not substance)
|
|
const emojiCount = (
|
|
trimmed.match(/[\u{1F300}-\u{1F9FF}\u{2600}-\u{26FF}\u{2700}-\u{27BF}\u{1FA00}-\u{1FAFF}]/gu) ||
|
|
[]
|
|
).length;
|
|
if (emojiCount > 3) {
|
|
return true;
|
|
}
|
|
|
|
return false;
|
|
}
|
|
|
|
export function passesAttentionGate(text: string): boolean {
|
|
const trimmed = text.trim();
|
|
|
|
// Length bounds
|
|
if (trimmed.length < MIN_CAPTURE_CHARS || trimmed.length > MAX_CAPTURE_CHARS) {
|
|
return false;
|
|
}
|
|
|
|
// Word count — short phrases ("I need those") lack context for recall
|
|
const wordCount = trimmed.split(/\s+/).length;
|
|
if (wordCount < MIN_WORD_COUNT) {
|
|
return false;
|
|
}
|
|
|
|
if (failsSharedGateChecks(trimmed)) {
|
|
return false;
|
|
}
|
|
|
|
// Passes gate — retain for short-term storage
|
|
return true;
|
|
}
|
|
|
|
// ============================================================================
|
|
// Assistant attention gate — stricter filter for assistant messages
|
|
// ============================================================================
|
|
|
|
/** Maximum assistant message length — shorter than user to avoid code dumps. */
|
|
const MAX_ASSISTANT_CAPTURE_CHARS = 1000;
|
|
|
|
/** Minimum word count for assistant messages — higher than user. */
|
|
const MIN_ASSISTANT_WORD_COUNT = 10;
|
|
|
|
/**
|
|
* Patterns that reject assistant self-narration — play-by-play commentary
|
|
* that reads like thinking out loud rather than a conclusion or fact.
|
|
* These are the single biggest source of noise in auto-captured assistant memories.
|
|
*/
|
|
const ASSISTANT_NARRATION_PATTERNS = [
|
|
// "Let me ..." / "Now let me ..." / "I'll ..." action narration
|
|
/^(ok[,.]?\s+)?(now\s+)?let me\s+(check|look|see|try|run|start|test|read|update|verify|fix|search|process|create|build|set up|examine|investigate|query|fetch|pull|scan|clean|install|download|configure|make|select|click|type|fill|open|close|switch|send|post|submit|edit|change|add|remove|write|save|upload)/i,
|
|
// "I'll ..." action narration
|
|
/^I('ll| will)\s+(check|look|see|try|run|start|test|read|update|verify|fix|search|process|create|build|set up|examine|investigate|query|fetch|pull|scan|clean|install|download|configure|execute|help|handle|make|select|click|type|fill|open|close|switch|send|post|submit|edit|change|add|remove|write|save|upload|use|grab|get|do)/i,
|
|
// "Starting ..." / "Running ..." / "Processing ..." status updates
|
|
/^(starting|running|processing|checking|fetching|scanning|building|installing|downloading|configuring|executing|loading|updating|filling|selecting|clicking|typing|opening|closing|switching|navigating|uploading|saving|sending|posting|submitting)\s/i,
|
|
// "Good!" / "Great!" / "Perfect!" / "Done!" as opener followed by narration
|
|
/^(good|great|perfect|nice|excellent|awesome|done)[!.]?\s+(i |the |now |let |we |that |here)/i,
|
|
// Progress narration: "Now I have..." / "Now I can see..." / "Now let me..."
|
|
/^now\s+(i\s+(have|can|need|see|understand)|we\s+(have|can|need)|the\s|on\s)/i,
|
|
// Step narration: "Step 1:" / "**Step 1:**"
|
|
/^\*?\*?step\s+\d/i,
|
|
// Page/section progress narration: "Page 1 done!", "Page 3 — final page!"
|
|
/^Page\s+\d/i,
|
|
// Narration of what was found/done: "Found it." / "Found X." / "I see — ..."
|
|
/^(found it|found the|i see\s*[—–-])/i,
|
|
// Sub-agent task descriptions (workflow narration)
|
|
/^\[?(mon|tue|wed|thu|fri|sat|sun)\s+\d{4}-\d{2}-\d{2}/i,
|
|
// Context compaction self-announcements
|
|
/^🔄\s*\*?\*?context reset/i,
|
|
// Filename slug generation prompts (internal tool use)
|
|
/^based on this conversation,?\s*generate a short/i,
|
|
|
|
// --- Conversational filler responses (not knowledge) ---
|
|
// "I'm here" / "I am here" filler: "I'm here to help", "I am here and listening", etc.
|
|
/^I('m| am) here\b/i,
|
|
// Ready-state: "Sure, (just) tell me what you want..."
|
|
/^Sure[,!.]?\s+(just\s+)?(tell|let)\s+me/i,
|
|
// Observational UI narration: "I can see the picker", "I can see the button"
|
|
/^I can see\s/i,
|
|
// A sub-agent task report (quoted or inline)
|
|
/^A sub-?agent task\b/i,
|
|
|
|
// --- Injected system/voice context (not user knowledge) ---
|
|
// Voice mode formatting instructions injected into sessions
|
|
/^\[VOICE\s*(MODE|OUTPUT)/i,
|
|
/^\[voice[-\s]?context\]/i,
|
|
// Voice tag prefix
|
|
/^\[voice\]\s/i,
|
|
|
|
// --- Session completion summaries (ephemeral, not long-term knowledge) ---
|
|
// "Done ✅ ..." completion messages (assistant summarizing what it just did)
|
|
/^Done\s*[✅✓☑️]\s/i,
|
|
// "All good" / "All set" wrap-ups
|
|
/^All (good|set|done)[!.]/i,
|
|
// "Here's what changed" / "Summary of changes" (session-specific)
|
|
/^(here'?s\s+(what|the|a)\s+(changed?|summary|breakdown|recap))/i,
|
|
|
|
// --- Open proposals / action items (cause rogue actions when recalled) ---
|
|
// These are dangerous in memory: when auto-recalled, other sessions interpret
|
|
// them as active instructions and attempt to carry them out.
|
|
// "Want me to...?" / "Should I...?" / "Shall I...?" / "Would you like me to...?"
|
|
/want me to\s.+\?/i,
|
|
/should I\s.+\?/i,
|
|
/shall I\s.+\?/i,
|
|
/would you like me to\s.+\?/i,
|
|
// "Do you want me to...?"
|
|
/do you want me to\s.+\?/i,
|
|
// "Can I...?" / "May I...?" assistant proposals
|
|
/^(can|may) I\s.+\?/i,
|
|
// "Ready to...?" / "Proceed with...?"
|
|
/ready to\s.+\?/i,
|
|
/proceed with\s.+\?/i,
|
|
];
|
|
|
|
export function passesAssistantAttentionGate(text: string): boolean {
|
|
const trimmed = text.trim();
|
|
|
|
// Length bounds (stricter than user)
|
|
if (trimmed.length < MIN_CAPTURE_CHARS || trimmed.length > MAX_ASSISTANT_CAPTURE_CHARS) {
|
|
return false;
|
|
}
|
|
|
|
// Word count — higher threshold than user messages
|
|
const wordCount = trimmed.split(/\s+/).length;
|
|
if (wordCount < MIN_ASSISTANT_WORD_COUNT) {
|
|
return false;
|
|
}
|
|
|
|
// Reject messages that are mostly code (>50% inside triple-backtick fences)
|
|
const codeBlockRegex = /```[\s\S]*?```/g;
|
|
let codeChars = 0;
|
|
let match: RegExpExecArray | null;
|
|
while ((match = codeBlockRegex.exec(trimmed)) !== null) {
|
|
codeChars += match[0].length;
|
|
}
|
|
if (codeChars > trimmed.length * 0.5) {
|
|
return false;
|
|
}
|
|
|
|
// Reject messages that are mostly tool output
|
|
if (
|
|
trimmed.includes("<tool_result>") ||
|
|
trimmed.includes("<tool_use>") ||
|
|
trimmed.includes("<function_call>")
|
|
) {
|
|
return false;
|
|
}
|
|
|
|
if (failsSharedGateChecks(trimmed)) {
|
|
return false;
|
|
}
|
|
|
|
// Assistant-specific narration patterns (play-by-play self-talk)
|
|
if (ASSISTANT_NARRATION_PATTERNS.some((r) => r.test(trimmed))) {
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|