Memory — Autonomous preference management

This skill implements Gold Rule 3 (user preferences always up to date). The agent proactively detects, stores, and injects user preferences without being asked.

When it runs

Always. This is not a flow triggered by a keyword. It is a background behavior that runs during every interaction:

1. Detection: after every user message, evaluate whether a preference was stated, implied, or corrected
2. Injection: at the pre-flight of every flow, load active preferences into context

Detection

After every user message, scan for 3 signal types:

Signal	Example	Confidence	Source
Explicit	"no quiero empresas de crypto"	1.0	explicit_statement
Implicit	"respondeme corto"	0.7	inferred
Correction	"actualmente busco roles de IC, no manager"	1.0	correction

Save vs skip checklist

Save (proactively, no need to ask):

• Job search preferences (roles, industries, locations, work mode, salary)
• Communication preferences (language, tone, length, format)
• Compensation criteria (range, equity, benefits)
• Tooling/workflow preferences (which platforms, how to apply)
• Corrections to anything previously stored
• Explicit requests: "recordá que..." / "remember that..."
• Strategy level changes. When the user's situation changes (employment status, urgency), detect and propose a strategy level change. See AGENTS.md "Strategy levels"

Skip:

• Trivial/obvious info ("user asked about Python")
• Already in CV or users.data.profile (don't duplicate)
• Already in users.data.style_profile
• Re-discoverable facts (can web search)
• Session-specific ephemera (temporary file paths, one-off debugging)
• Already in AGENTS.md or other context files

Storage

All preferences live in the preferences table, accessed via scripts/db.js:

preferences (
  id SERIAL PRIMARY KEY,
  user_id INTEGER REFERENCES users(id),
  category TEXT NOT NULL,        -- job_search, communication, compensation, tooling, workflow
  key TEXT NOT NULL,             -- e.g. "avoid_industries", "reply_language", "salary_min"
  value TEXT NOT NULL,           -- e.g. "crypto,gambling", "spanish", "5000"
  confidence REAL DEFAULT 1.0,   -- 1.0 explicit, 0.7 inferred, 0.5 auto-summarized
  source TEXT DEFAULT 'explicit_statement', -- explicit_statement | inferred | correction
  status TEXT DEFAULT 'active',  -- active | superseded
  created_at TIMESTAMPTZ DEFAULT NOW(),
  updated_at TIMESTAMPTZ DEFAULT NOW(),
  UNIQUE(user_id, category, key)
)

Categories

Category	Examples
job_search	avoid_industries, target_roles, location_preference, work_mode, visa_requirements
communication	reply_language, reply_tone, reply_length, greeting_style, avoid_bullets
compensation	salary_min, salary_max, currency, equity_expectation, required_benefits
tooling	preferred_platforms, apply_method, browser_mode
workflow	application_batch_size, follow_up_timing, auto_apply_threshold, strategy_level

Save a new preference

node scripts/db.js "INSERT INTO preferences (user_id, category, key, value, confidence, source) VALUES (<user_id>, '<category>', '<key>', '<value>', <confidence>, '<source>') ON CONFLICT (user_id, category, key) DO UPDATE SET value = EXCLUDED.value, confidence = EXCLUDED.confidence, source = EXCLUDED.source, updated_at = NOW()" --write

The ON CONFLICT clause handles updates: if the preference already exists, it replaces the value and bumps updated_at. No need to check first.

Correction handling

When a correction is detected (user contradicts a stored preference), the same INSERT with ON CONFLICT DO UPDATE handles it. The old value is replaced, source becomes correction, and updated_at is bumped. The old value is not preserved (single user, no audit trail needed).

Strategy level detection

The agent must detect signals that the user's job search situation has changed and propose a strategy level adjustment. This is Gold Rule 3 applied to urgency/aggressiveness.

Signal	Example	Proposed level
Lost job / fired	"me despidieron", "me quedé sin trabajo", "lost my job"	active
About to lose job	"me van a despedir", "termina mi contrato en X", "my contract ends"	active
Desperation	"necesito algo ya", "urgentísimo", "need a job now"	aggressive
Found a job	"encontré trabajo", "acepté una oferta", "got the job"	passive
Employed, casually looking	"estoy viendo opciones", "open to opportunities"	selective
Wants more aggressive	"aplica más agresivo", "send more applications"	bump up one level
Wants less aggressive	"frená un poco", "too many applications"	bump down one level

When a strategy change is detected:

1. Propose the change to the user (never change without asking)
2. Explain what will change (batch sizes, match threshold, relax rules)
3. Wait for confirmation
4. Save to preferences (workflow.strategy_level) and users.data.strategy
5. Report: "Estrategia actualizada: <level>"

Injection

At the pre-flight of every flow, load active preferences:

node scripts/db.js "SELECT category, key, value, confidence, source FROM preferences WHERE user_id = <user_id> AND status = 'active' ORDER BY category, key"

Inject the result into the flow's context. Treat preferences as constraints:

• job_search.* → filter jobs, discard non-matching
• communication.* → shape drafts (language, tone, length)
• compensation.* → filter by salary, negotiate
• tooling.* → choose platforms and methods
• tooling.browser_mode → controls browser visibility in all flows that use the browser (agent-browser). The preference maps to agent-browser.json headed / --headed/--headless flags. Valid values: headless (always headless except manual login/2FA), headed (always headed), headed_logins_only (headed only for logins/2FA, headless otherwise), ask_each_time (agent asks before each browser session). Default if not set: headed_logins_only. Manual login/2FA is always headed regardless of this preference (Gold Rule 5)
• workflow.* → tune batch sizes and timing

Confidence-aware: preferences with confidence < 1.0 (inferred) can be overridden by explicit user instructions in the current session. Preferences with confidence = 1.0 (explicit) hold unless the user explicitly changes them.

Rules

• Proactive, not reactive. Save when detected, don't wait to be asked
• All DB access via scripts/db.js (see db skill)
• Don't duplicate. If a preference is already in users.data.profile or users.data.style_profile, skip it. The preferences table is for things said in conversation that aren't captured by the structured profile
• Single source of truth per (category, key). Use ON CONFLICT DO UPDATE, never insert duplicates
• Corrections replace, not append. "actually I want X" updates the existing value
• Inferred preferences are weaker. Mark confidence = 0.7 and source = 'inferred'. They can be overridden by explicit statements
• Never ask "should I save this?". If it passes the save checklist, save it silently
• Report what was saved. After saving, briefly mention: "Guardé que preferís X" / "Saved preference: X". One line, no ceremony

Dependencies

• Depends on onboarding (DB + users table must exist)
• Uses db skill for all DB access
• Consumed by every flow via pre-flight injection