Skip to main content

​
Overview

Memories are Sphinx’s semantic cache, a persistent layer that captures and recalls the knowledge that typically lives in your head or gets lost between sessions. Unlike traditional context windows that reset with each conversation, memories build up over time to create an understanding of your work and data. Think of memories as the institutional knowledge that usually only exists as spoken history: the quirks of your datasets, the naming conventions your team prefers, the relationships between tables that aren’t documented anywhere, and the preferences you’ve developed through experience.

​
What Memories Capture

Sphinx automatically observes and stores several categories of knowledge:

​
User Preferences

Sphinx learns how you like to work:
  • Preferred coding styles and formatting conventions
  • Visualization choices (color schemes, chart types, libraries)
  • Column naming conventions and data transformation patterns
  • How you prefer explanations (brief vs. detailed, technical vs. accessible)

​
Data Relationships

As you work with your data, Sphinx builds an understanding of:
  • How tables and datasets relate to each other
  • Key columns and their semantic meaning
  • Common joins and aggregation patterns
  • Data quality issues and known edge cases

​
Siloed Knowledge

The undocumented knowledge that makes you effective:
  • Business logic that isn’t captured in code comments
  • Historical context about why certain decisions were made
  • Domain-specific terminology and abbreviations
  • Gotchas and pitfalls learned from experience
Memories are especially powerful for recurring analyses. Information you share once—like “revenue numbers should always exclude refunds” or “the status column uses legacy codes from the old system”—gets remembered and applied automatically in future sessions.

​
How Memories Work

Memories operate as a semantic cache, meaning Sphinx retrieves relevant memories based on semantic meaning rather than exact keyword matches. When you start working on a task, Sphinx searches its memory store for contextually relevant information and surfaces it automatically. This creates a virtuous cycle: the more you use Sphinx, the more it learns about your specific context, and the more relevant its assistance becomes.

​
Managing Memories

​
Viewing Memories

Access your stored memories through the Sphinx Memory interface inside the extension to see what observations have been captured. This helps you understand what Sphinx “knows” about your work.

​
Creating Memories Manually

You can explicitly tell Sphinx to remember something:
“Remember that the customer_id field in the orders table corresponds to id in the legacy_customers table, not the new customers table.”
“Remember that Q4 reports should always use fiscal year dates, not calendar year.”

​
Updating and Removing Memories

If a memory becomes outdated or incorrect, you can:
  • Ask Sphinx to update a specific memory with new information
  • Delete memories that are no longer relevant
  • Clarify or correct misunderstandings that were captured
Memories persist across sessions and workspaces. Be mindful about what information you ask Sphinx to remember, especially when working with sensitive data or in shared environments.

​
Best Practices

When you encounter information that will be useful across sessions, tell Sphinx directly. Explicit memories are more reliable than hoping Sphinx will infer the right context.
If Sphinx makes an incorrect assumption based on a memory, correct it promptly. This updates the memory and prevents the same mistake from recurring.
Memories work best for information that doesn’t change frequently—data relationships, business rules, and preferences. For rapidly changing context, rely on in-session conversation instead.
As your data and workflows evolve, some memories may become stale. Periodic review helps keep Sphinx’s understanding current and accurate.

​
Example Use Cases

ScenarioMemory Example
Data quirks”The timestamp column in events is stored in UTC but the dashboard expects Pacific time”
Team conventions”We use snake_case for all DataFrame columns”
Business logic”Churn is defined as no activity in 90 days, not account cancellation”
Historical context”The 2023 Q2 data has known gaps due to the migration”
Personal preference”Always show percentages with one decimal place”