Module memory_compaction 

Hierarchical Memory Compaction (Task 5)

This module implements semantic memory compaction inspired by LSM-trees. It manages tiered storage where older memories are summarized to maintain bounded context while preserving semantic continuity.

Architecture

L0: Raw Episodes (recent, full detail)
    │
    ▼ Summarization
L1: Summaries (older, compressed)
    │
    ▼ Abstraction
L2: Abstractions (oldest, highly compressed)

Compaction Strategy

• Episodes older than tier threshold are grouped by semantic similarity
• Each group is summarized (via LLM or extractive methods)
• Summaries are re-embedded for retrieval
• Growth is O(log_c T) where c = compaction ratio, T = total events

Structs

Abstraction

An abstraction (L2)

CompactionStats

Compaction statistics

Episode

A raw episode (L0)

ExtractiveSummarizer

Extractive summarizer (no LLM required)

HierarchicalMemory

Hierarchical memory store with compaction

MemoryCompactionConfig

Configuration for memory compaction

Summary

A summary (L1)

Enums

CompactionError

Compaction error

EpisodeType

Episode types

MemoryEntry

Entry from hierarchical memory

Traits

Summarizer

Trait for summarization backends

Functions

create_hierarchical_memory

Create a hierarchical memory with extractive summarizer

create_test_memory

Create with aggressive compaction for testing