YourMemory Launches AI Memory System Mimicking Human Forgetting—Achieves 59% Recall Rate

New Tool Gives AI Agents Persistent Memory That Decays Like Human Brains

A developer has unveiled YourMemory, a memory management system designed to equip artificial intelligence agents with persistent recall capabilities that function similarly to human memory. Unlike conventional AI assistants that reset after each session, YourMemory enables agents to retain user preferences, technical decisions, and contextual information across multiple conversations while automatically deprioritizing outdated or less critical data.

The system was benchmarked against competitor Zep Cloud on the LoCoMo-10 dataset, comprising 1,534 question-answer pairs from 10 multi-session conversations. YourMemory achieved a 59% recall rate at the fifth position (with a 95% confidence interval of 56–61%), more than double Zep Cloud's 28% recall rate (26–30% CI).

How the System Works

YourMemory applies principles from cognitive psychology, specifically the Ebbinghaus forgetting curve, which models how human memory strength diminishes over time. The system uses a mathematical formula that factors in importance weighting and recall frequency to slow memory decay:

effective_λ = base_λ × (1 - importance × 0.8)
strength = importance × e^(−effective_λ × days) × (1 + recall_count × 0.2)
score = cosine_similarity × strength

Memories are categorized into four types, each with different retention periods: strategy (approximately 38 days without recall), fact (approximately 24 days), assumption (approximately 19 days), and failure (approximately 11 days). Memories below a strength threshold of 0.05 are automatically pruned every 24 hours.

The retrieval system operates in two rounds. Round 1 performs vector similarity search across all stored memories. Round 2 uses graph-based expansion via breadth-first search to surface contextually related memories that may not share direct vocabulary overlap, identifying connections through semantic edges with similarity scores of 0.4 or higher.

Installation and Integration

The tool requires minimal setup. Installation demands Python versions 3.11, 3.12, 3.13, or 3.14 and involves two commands:

• pip install yourmemory
• yourmemory-setup

No Docker containers, external database services, or infrastructure configuration is necessary. The system stores data locally in a DuckDB database at ~/.yourmemory/memories.duckdb and integrates with multiple AI development platforms through MCP (Model Context Protocol) server standards.

Supported platforms include Claude Code, Claude Desktop, Cline (VS Code extension), Cursor, and OpenCode. Integration involves adding a configuration block to the platform's settings file and restarting the application. After integration, developers add a workflow instruction file (sample_CLAUDE.md) to their project directory.

Core Functionality

The system provides three primary tools that Claude and other AI agents call automatically:

• recall_memory(query) — Activates at the beginning of each task, surfacing relevant memories ranked by similarity and strength metrics
• store_memory(content, importance) — Triggered after the AI learns new information, embedding and storing data with biological decay characteristics
• update_memory(id, new_content) — Replaces outdated memories by re-embedding and refreshing the stored entry

In a practical example, if a developer stores the fact "Sachit prefers tabs over spaces in Python" with an importance rating of 0.9, the system will retain this preference across sessions. When queried about Python formatting in a later conversation without explicit reminder, the AI recalls the stored preference with a strength value of 0.87.

Multi-Agent and Enterprise Capabilities

Multiple agents can operate within a single YourMemory instance while maintaining isolated private memory stores and controlled shared context access. Developers register agents using an agent registration function and receive unique API keys for authentication. Access permissions are configurable, allowing agents to read or write to shared and private memory categories independently.

For larger deployments, the system supports optional PostgreSQL with pgvector or Neo4j graph database backends beyond the default DuckDB and NetworkX combination.

Technical Architecture

The default technology stack includes DuckDB for vector storage with native cosine similarity calculations, NetworkX for graph operations, sentence-transformers for local embeddings (768-dimensional all-mpnet-base-v2 model), spaCy for natural language processing tasks including deduplication, and APScheduler for automated 24-hour decay jobs.

The system references the LoCoMo dataset published by Snap Research in 2024 for benchmarking purposes. The source code is distributed under a CC-BY-NC-4.0 license, permitting personal, educational, academic, and open-source use, with commercial licensing available through separate agreement.