Memory System Overview

The Context Graph memory system provides long-term memory management for AI agents, including memory consolidation (summarizing and compacting old memories), efficient retrieval (combining recency, relevance, and entity proximity), and context deduplication (removing redundant information before LLM context assembly).

Architecture

+------------------+     +--------------------+     +-------------------+
| Temporal Memory  | --> | Memory Retrieval   | --> | Context Assembly  |
| Store            |     | Service            |     | (for LLM)        |
+------------------+     +--------------------+     +-------------------+
        |                         |
        v                         v
+------------------+     +--------------------+
| Memory           |     | Context            |
| Consolidation    |     | Deduplication      |
+------------------+     +--------------------+

Subsections

Page	Description
Memory Consolidation	LLM-powered summarization and forgetting policies
Memory Retrieval	Multi-factor memory recall combining recency, relevance, and proximity
Context Deduplication	Removing duplicate contexts before LLM input assembly

Memory Fact Types

Type	Description	Example
`PREFERENCE`	User or agent preference	"User prefers bar charts for time series"
`LEARNED_FACT`	A fact learned from interaction	"Sales data is refreshed daily at 6am"
`CONTEXT_NOTE`	Contextual information	"Q4 includes holiday season adjustments"
`CORRECTION`	A correction to previous output	"Revenue should exclude returns"
`SCHEMA_NOTE`	Schema-related knowledge	"customer_id is the primary key in orders table"

Memory Importance Levels

Level	Description	Retention Priority
`CRITICAL`	Must never be forgotten	Highest
`HIGH`	Important for task quality	High
`MEDIUM`	Useful but not essential	Medium
`LOW`	Nice to have	Low
`EPHEMERAL`	Temporary, can be discarded	Lowest

Feedback Scoring Memory Consolidation