Zep provides long-term memory for AI assistants. pref0 provides preference learning. Both improve personalization, but they approach it differently.
| pref0 | Zep | |
|---|---|---|
| What it stores | Structured preferences with confidence scores | Conversation history, summaries, and facts |
| How it learns | Extracts preferences from corrections | Summarizes and indexes conversations |
| Confidence over time | Yes — preferences compound across sessions | No — facts are stored without confidence |
| Retrieval | Full preference profile per user | Semantic search over memory |
| Infrastructure | Hosted API, no setup | Self-hosted or cloud, requires more setup |
| Best for | Learning preferences from corrections | Long-term conversation memory and recall |
pref0 extracts structured key-value preferences with confidence scores. Zep stores unstructured conversation history and generates summaries. pref0's structured format is easier to inject into system prompts.
pref0 actively learns from corrections and compounds confidence. Zep remembers what was said. The distinction matters: 'use TypeScript' said three times becomes a high-confidence preference in pref0, but just three separate memories in Zep.
pref0 is 2 endpoints. Zep is a full memory platform with conversation management, fact extraction, and semantic search. If you only need preferences, pref0 is significantly simpler.
Yes. Use Zep for conversation memory and context retrieval. Use pref0 specifically for preference learning with confidence scoring. They solve different problems.
No. pref0 processes conversations to extract preferences, then discards the raw conversation data. If you need conversation storage, use Zep alongside pref0.
pref0 is a hosted API with no infrastructure requirements. Zep can be self-hosted or cloud-hosted, with more setup involved.
Your users are already teaching your agent what they want. pref0 makes sure the lesson sticks.