YourMemory: AI memory with biological decay hits 59% recall on LoCoMo-10

YourMemory implements persistent memory for AI agents using the Ebbinghaus forgetting curve — memories decay unless reinforced by recall, and unused data is pruned when it hits a threshold. Built as a local-first MCP server on DuckDB, it combines BM25, vector search, and a graph layer to solve the "logical neighbor" problem where semantic search misses relevant but non-similar nodes.

Benchmarks

On the LoCoMo-10 benchmark (1,534 QA pairs across 10 multi-session conversations):

• YourMemory: 59% Recall@5 (95% CI: 56–61%)
• Zep Cloud: 28% (95% CI: 26–30%)

That's 2× better recall than Zep Cloud. Stateless vector stores reportedly suffer 84% more token waste.

Quick Start

Python 3.11–3.14. No Docker or external services needed.

pip install yourmemory
yourmemory-setup

Get your config path:

yourmemory-path

MCP Configuration

Claude Code — add to ~/.claude/settings.json:

{
  "mcpServers": {
    "yourmemory": {
      "command": "yourmemory"
    }
  }
}

Claude Desktop — add to the appropriate config file:

{
  "mcpServers": {
    "yourmemory": {
      "command": "yourmemory"
    }
  }
}

Cline, Cursor, OpenCode, and any MCP-compatible client (Windsurf, Continue, Zed) can wire it in using the full path from yourmemory-path.

Memory Workflow

Copy the sample instructions:

cp sample_CLAUDE.md CLAUDE.md

Then edit CLAUDE.md with your name and user ID. Claude follows a recall → store → update workflow on every task using three MCP tools:

• recall_memory(query) — surfaces relevant memories at start of task
• store_memory(content, importance) — embeds and stores with biological decay
• update_memory(id, new_content) — re-embeds and replaces outdated info

Example: store_memory("Sachit prefers tabs over spaces in Python", importance=0.9, category="fact")

Who It's For

Developers building AI coding agents that run long-lived projects and need to remember user preferences, project context, and avoid retraining from scratch each session.