Memory

Kaizen has a persistent memory system that improves its understanding of you and your preferences over time.

User Memory

User memory is a global profile that the agent maintains across all conversations. It stores things like:

  • Your preferences and how you like things done

  • Accounts and services you use

  • Past decisions and their context

  • Patterns the agent has noticed

Viewing & Editing

Go to Settings > Memory (/settings/memory) to view and edit the user memory directly. It's a plain text field that auto-saves when you click away.

The agent can also read and write to this memory during runs using the read-user-memory and write-user-memory tools.

Memory Compaction

When the agent writes to user memory, it doesn't just append. It uses an LLM-based merge process to deduplicate, consolidate, and compress the information. This keeps the memory concise and relevant as it grows.

Working Memory

Working memory is per-objective context that helps the agent across multiple runs on the same task. Unlike user memory (which is global), working memory is scoped to a specific objective and captures:

  • What was tried in previous runs

  • What worked and what didn't

  • Context that should carry forward

Tool Memory

Individual tools can have their own persistent memory. When the agent learns something useful about a tool (quirks, tips, common patterns), it saves it to the tool's memory field. This knowledge is available in future runs.

Learning Loop

The memory system improves with use:

  1. A run completes

  2. Insights and learnings from the interaction are persisted to user memory

  3. Future runs benefit from these accumulated insights

This is what makes Kaizen's "self-improving" claim concrete. The more you use it, the better it gets at understanding your needs and avoiding past mistakes.

Last updated