Why don't user profile preferences decay? I currently have 100% satisfaction for each one

[haisto]

opencode-mem: User Profile Confidence Decay Analysis

Plugin: opencode-mem v2.14.2
Analyzed: May 28, 2026

---

1. Overview

opencode-mem builds a User Profile by periodically analyzing user prompts with an AI model (Claude/GPT). The profile includes preferences (code style, communication style, tool preferences), each with a confidence score (0–1) meant to represent how certain the system is about that preference.

However, due to a chain of implementation issues, confidence never decays — it only increases until reaching 100%, where it stays forever.

---

2. Core Types

interface UserProfilePreference {
    category: string;       // e.g. "Code Style", "Communication"
    description: string;    // e.g. "Prefers concise code without comments"
    confidence: number;     // 0–1, meant to represent certainty
    evidence: string[];     // Example prompts supporting this preference
    lastUpdated: number;    // Timestamp of last confirmation
}

interface UserProfileData {
    preferences: UserProfilePreference[];
    patterns: UserProfilePattern[];
    workflows: UserProfileWorkflow[];
}

---

3. The Flow

User sends prompts
        │
        ▼
┌─────────────────────────────┐
│  UserPromptManager          │  saves every prompt to DB
│  (user-prompt-manager.js)   │
└─────────────┬───────────────┘
              │
              ▼  (every N prompts where N = userProfileAnalysisInterval, default 10)
┌─────────────────────────────┐
│  performUserProfileLearning │  calls AI model to extract profile data
│  (user-memory-learning.js)  │
└─────────────┬───────────────┘
              │
              ▼
┌─────────────────────────────┐
│  mergeProfileData()         │  merges AI output with existing profile
│  (user-profile-manager.js)  │
└─────────────┬───────────────┘
              │
              ▼
     ┌ ─ ─ ─ ─ ─ ─ ─ ┐
     │ applyConfidenceDecay() │  ← DEAD CODE, never called
     └ ─ ─ ─ ─ ─ ─ ─ ┘

---

4. The Three Broken Links

Link ① — Merge Only Increases, Never Decreases

File: user-profile-manager.js, line 218

// Inside mergeProfileData(), when a matching preference already exists:
merged.preferences[existingIndex] = {
    ...newPref,
    confidence: Math.min(1, (existingItem.confidence || 0) + 0.1),
    //                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    //                     Always adds +0.1, capped at 1.0.
    //                     No branch for decreasing confidence.
    evidence: [
        ...new Set([
            ...this.ensureArray(existingItem.evidence),
            ...this.ensureArray(newPref.evidence),
        ]),
    ].slice(0, 5),
    lastUpdated: Date.now(),
};

Effect: Every time the same preference is extracted again by the AI, its confidence gets +0.1. After 2–3 analysis cycles, it hits 1.0 and stays there.

Link ② — Decay Function Exists But Is Never Called

File: user-profile-manager.js, lines 138–160

applyConfidenceDecay(profileId) {
    const profile = this.getProfileById(profileId);
    if (!profile) return;

    const profileData = JSON.parse(profile.profileData);
    const now = Date.now();
    const decayThreshold = CONFIG.userProfileConfidenceDecayDays * 24 * 60 * 60 * 1000;
    //                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    //                       Default: 30 days

    profileData.preferences = profileData.preferences
        .map((pref) => {
            const age = now - pref.lastUpdated;
            if (age > decayThreshold) {
                const decayFactor = Math.max(0.5, 1 - (age - decayThreshold) / decayThreshold);
                return { ...pref, confidence: pref.confidence * decayFactor };
            }
            return pref;
        })
        .filter((pref) => pref.confidence >= 0.3);
    // ...
}

A grep for applyConfidenceDecay across the entire codebase yields only one match — the definition itself. No caller exists anywhere.

Search scope	Matches
Definition in user-profile-manager.js	✅ 1 (line 138)
Callers in any .js file	❌ 0

Link ③ — Refresh API Is a Shell

File: api-handlers.js, lines 767–790

export async function handleRefreshProfile(userId) {
    // Step 1: Count unanalyzed prompts
    const unanalyzedCount = userPromptManager.countUnanalyzedForUserLearning();

    // Step 2: Return "queued" — does nothing else
    return {
        success: true,
        data: {
            message: "Profile refresh queued",
            unanalyzedPrompts: unanalyzedCount,
            note: "Profile will be updated when threshold is reached",
            //       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
            //       No actual refresh happens. No decay. No re-analysis.
        },
    };
}

The endpoint POST /api/user-profile/refresh is misleading — it only reports how many prompts are pending. It should, at minimum, call applyConfidenceDecay before returning.

Relevant config (opencode-mem.jsonc):

"userProfileConfidenceDecayDays": 30,
"userProfileAnalysisInterval": 10,
"userProfileMaxPreferences": 20,

Despite the config existing, userProfileConfidenceDecayDays is effectively useless because the only function that reads it (applyConfidenceDecay) is never invoked.

---

5. Root Cause Summary

#	Issue	Location	Impact
①	mergeProfileData() hardcodes +0.1 on every merge	user-profile-manager.js:218	Confidence can only increase
②	applyConfidenceDecay() is dead code — defined but never called	user-profile-manager.js:138	Decay mechanism is non-functional
③	POST /api/user-profile/refresh is a no-op shell	api-handlers.js:767	No manual trigger for decay exists

---

6. Ideal Fix (Reference)

An upstream fix would need to address all three links:

1. Add a configurable merge strategy (e.g., confidenceMergeStrategy: "max" | "increment" | "ai-score")
2. Hook applyConfidenceDecay() into the profile analysis pipeline
3. Make the refresh API actually trigger decay + re-analysis

Until then, confidence will always reach and stay at 100% for any preference that appears more than once.