agent user learning

Maintaining a consistent persona across sessions often fails because LLMs forget previous preferences or bury them in long chat histories. This skill implements Agent User Learning (AUL) to solve the "memory" problem for developers building persistent agents. It provides a structured protocol to store, retrieve, and update user-specific attributes and behavior rules without bloating the context window with raw transcripts.

What it does

• Persistent memory manages a private database of user-specific attributes, interests, and stated goals.
• Dynamic behavior adaptation adjusts agent responses based on stored user preferences for tone, depth, and workflow.
• Scoped context retrieval pulls only relevant knowledge (e.g., coding or writing) to keep tokens low and focus high.
• Explicit correction handling allows users to update or delete stored information, ensuring the agent stays aligned with current needs.
• Corroborated inference tracks patterns across sessions, only promoting inferred preferences to stable state after multiple confirmations.

How it works

1. Initialize the agent by resolving the user's private database using a consistent --user-id.
2. Retrieve Context by selecting specific scopes relevant to the current task to load into the prompt.
3. Execute Task using the loaded attributes to tailor examples, starting levels, and communication styles.
4. Observe & Learn by extracting useful summaries from the interaction and committing them back to the database as idempotent events.

Frameworks & tools

The skill uses a Python-based CLI tool (aul.py) to interface with a local or remote SQLite database. It is designed to be integrated into custom agent lifecycles or wrappers around OpenAI, Anthropic, or local LLM APIs.

Why this beats prompting it yourself

Relying on a system prompt to "remember everything" leads to context overflow and hallucinated preferences. This skill provides a rigorous, versioned event ledger that distinguishes between hypothetical discussion and actual user evidence, preventing the agent from becoming a caricatured version of the user.

Use cases

• Coding assistants that remember your specific library preferences and documentation style across different repositories.
• Content creation agents that maintain a consistent brand voice and avoid previously rejected phrasing.
• Personalized tutors that track a student's current knowledge level and preferred learning modalities.
• Workflow automation that remembers specific deployment constraints and notification preferences for a developer.

Known limitations

Does not include background workers or automatic hooks; the host application must trigger the read/write events. It stores summaries, not full conversation transcripts.