Module memory 

Memory system for MatrixCode.

This module implements automatic memory accumulation for AI agents. It captures user preferences, project decisions, key findings, and solutions across sessions, providing persistent context.

Module Structure

• config: Constants and configuration
• types: Core types (MemoryCategory, MemoryEntry, AutoMemory)
• storage: File storage with locking
• retrieval: TF-IDF search, keyword extraction
• extractor: AI and rule-based memory detection
• learning: Feedback learning, behavior inference
• project: Project structure analysis
• keywords_config: Customizable keywords configuration

Structs

AiMemoryExtractor

AI-based memory extractor using a fast/cheap model.

AutoMemory

Manager for automatic memory accumulation.

BehaviorInference

Result of behavior inference.

BehaviorInferenceConfig

Configuration for behavior inference.

FeedbackResult

Result of feedback detection.

KeywordsConfig

Keywords configuration for memory detection.

MemoryConfig

Configuration for the memory system.

MemoryEntry

A single memory entry.

MemoryFileLock

File lock for preventing concurrent access to memory storage.

MemoryStatistics

Statistics about memory collection.

MemoryStorage

Storage for memory files (global and project-level) with file locking.

ProjectStructureAnalyzer

ProjectTypeConfig

Project type detection configuration.

StopWordsConfig

Stop words configuration.

TfIdfSearch

Semantic search using TF-IDF algorithm.

Enums

AiDetectionMode

AI memory detection mode. Controls whether AI is used for memory category detection.

AiKeywordMode

AI keyword extraction mode.

FeedbackAction

Action to take when user feedback is detected.

MemoryCategory

Categories for memory entries.

Constants

CONFLICT_OVERLAY_THRESHOLD

Topic overlap threshold for conflict detection.

CONFLICT_OVERLAY_THRESHOLD_WITH_SIGNAL

Lower topic overlap threshold when change signal is present.

CONTEXT_IMPORTANCE_WEIGHT

Weight for importance in contextual summary (1.0 - CONTEXT_RELEVANCE_WEIGHT).

CONTEXT_RELEVANCE_WEIGHT

Weight for relevance in contextual summary (relevance vs importance trade-off).

DEFAULT_FAST_MODEL

Default fast model for AI memory extraction.

DEFAULT_IMPORTANCE_DECISION

Default importance scores by category. Lower values allow for gradual importance growth through references.

DEFAULT_IMPORTANCE_FINDING

DEFAULT_IMPORTANCE_PREF

DEFAULT_IMPORTANCE_SOLUTION

DEFAULT_IMPORTANCE_STRUCTURE

DEFAULT_IMPORTANCE_TECH

DEFAULT_MEMORY_EXTRACTOR_MODEL

Default model for cost-effective memory extraction.

IGNORE_DIRS

IMPORTANCE_STAR_THRESHOLD

Importance threshold for displaying star marker (⭐).

MAX_DETECTED_ENTRIES

Maximum entries to return from detection (to avoid overwhelming).

MAX_DISPLAY_LENGTH

Maximum length for display (shorter for terminal readability).

MAX_IMPORTANCE_CEILING

Maximum importance score ceiling (entries cannot exceed this).

MAX_MEMORY_CONTENT_LENGTH

Maximum length for memory content before truncation.

MERGE_SIMILARITY_THRESHOLD

Similarity threshold for merging similar memories (0.0-1.0). Lower than duplicate threshold to allow semantic merging.

MIN_KEYWORDS_FOR_AI_FALLBACK

Minimum keywords threshold for triggering AI fallback. If rule-based extraction produces fewer keywords than this, AI is used.

MIN_MEMORY_CONTENT_LENGTH

Minimum content length for memory detection (to avoid capturing too generic content). Increased to 20 to filter out short fragments.

MIN_SIMILARITY_LENGTH

Minimum content length for similarity check (to avoid short words matching everything).

PROJECT_TYPE_CONFIGS

SEMANTIC_ALIASES_DEFAULT

Default semantic aliases (embedded for fallback).

SIMILARITY_THRESHOLD

Similarity threshold for considering entries as duplicates (0.0-1.0). Higher value (0.85) reduces duplicate detection false negatives.

Traits

MemoryExtractor

Trait for memory extraction implementations.

Functions

apply_behavior_inferences_to_memory

Apply behavior inferences to memory.

apply_feedback_to_memory

Apply feedback to memory.

apply_tool_learning_to_memory

Generic tool learning - let model decide what to remember. This is a placeholder for future AI-driven learning.

calculate_similarity

Calculate word-based similarity between two strings (Jaccard coefficient).

compute_relevance

Compute relevance score of a memory entry to context keywords. Returns 0.0-1.0 where 1.0 means highly relevant.

detect_feedback_patterns

Detect user feedback patterns - generic detection, not exhaustive. The model should use its understanding to detect nuances.

detect_memories_fallback

Detect memories from text using configurable patterns.

detect_memories_from_text

Detect memories from text (wrapper for fallback).

detect_memories_smart

Smart detection: AI-first with rule-based fallback.

expand_semantic_keywords

Expand keywords with semantic aliases from KeywordsConfig.

extract_context_keywords

Extract meaningful keywords from conversation context. Uses KeywordsConfig for stop words and tech keywords. Improved: avoids meaningless character fragments.

extract_keywords_hybrid

Extract keywords using hybrid approach (rule-based + AI fallback).

generate_project_structure_memories

Generate project structure memories and save.

get_default_keywords

Get embedded default keywords configuration.

get_semantic_aliases

Get semantic aliases from KeywordsConfig.

has_contradiction_signal

Detect if two memory contents have contradiction signals. Uses KeywordsConfig for contradiction signals.

infer_category_from_content

Infer category from content.

infer_preferences_from_behavior

Infer patterns from behavior - generic word frequency analysis. Let the model decide what’s meaningful, not hardcoded tech patterns.

inference_to_memory_entry

Convert inference to memory entry.