Memory Consolidation

The MemoryConsolidationService manages long-term memory by summarizing old memories using LLM-powered summarization, scoring memory importance for retention priority, and applying forgetting policies to manage storage growth. It follows patterns from the Zep/Graphiti temporal memory architecture.

Overview

As agents accumulate memories over time, the raw memory store grows unbounded. Consolidation compacts this knowledge by grouping related memories, summarizing them into concise facts, and discarding low-importance memories according to configurable forgetting policies.

Source: data-plane/ai-service/src/context_graph/services/memory_consolidation_service.py

Consolidation Strategies

Strategy	Description	Best For
`SEMANTIC`	Group memories by semantic similarity	Related facts about the same topic
`TEMPORAL`	Group memories by time windows	Session-based consolidation
`ENTITY`	Group memories by linked entities	Entity-focused knowledge
`TYPE`	Group memories by fact type	Organizing preferences vs. learned facts
`HYBRID`	Combine multiple strategies	General-purpose consolidation

Forgetting Policies

Policy	Description
`AGE_BASED`	Forget memories older than a threshold
`IMPORTANCE_BASED`	Forget lowest importance memories first
`ACCESS_BASED`	Forget least recently accessed memories
`COMBINED`	Weighted combination of age, importance, and access
`ADAPTIVE`	Dynamically adjust based on storage pressure

Importance Scoring

Memory importance is computed from multiple factors:

Factor	Description	Weight
`BASE_IMPORTANCE`	Original importance level	0.30
`ACCESS_FREQUENCY`	How often the memory is accessed	0.20
`RECENCY`	How recently the memory was created	0.15
`ENTITY_LINKS`	Number of connected entities	0.15
`CONFIDENCE`	Fact confidence score	0.10
`FEEDBACK`	User feedback signals	0.10

Consolidation Process

Select -- Query memories eligible for consolidation based on age and status
Group -- Cluster memories using the configured strategy
Summarize -- Use an LLM to generate concise summaries for each group
Score -- Compute importance scores for the new consolidated memories
Replace -- Store consolidated memories and mark originals as archived
Forget -- Remove memories below the importance threshold

Configuration

from context_graph.services.memory_consolidation_service import (
    ConsolidationConfig,
    ConsolidationStrategy,
    ForgettingPolicy,
)
 
config = ConsolidationConfig(
    strategy=ConsolidationStrategy.HYBRID,
    forgetting_policy=ForgettingPolicy.COMBINED,
    max_memory_age_days=90,
    importance_threshold=0.3,
    max_memories_per_consolidation=100,
    target_compression_ratio=0.3,
)

Integration

The consolidation service integrates with:

Component	Purpose
`TemporalMemoryStore`	Source and destination for memory facts
`MemoryRetrievalService`	Provides access frequency data for importance scoring
LLM Provider	Generates summaries for memory groups
`BiTemporalStore`	Analytics queries for growth metrics

Memory System Overview Memory Retrieval