YourMemory Brings Biological Decay to AI Recall

A developer just shipped YourMemory, a persistent memory layer for AI agents that mimics how human brains hold onto important information and let unused facts fade. According to Hacker News, where the Show HN post climbed to a score of 175, YourMemory plugs into Claude Code, Claude Desktop, Cursor, Cline, and any MCP-compatible client through two install commands and zero infrastructure.

What stands out here is the benchmark. On LoCoMo-10 (1,534 QA pairs across 10 multi-session conversations), YourMemory hit 59% recall at top-5 results, versus 28% for Zep Cloud on the same test. That’s roughly 2x better recall against a well-known competitor.

How it works

1. Biological decay via the Ebbinghaus curve. Memory strength fades exponentially, but importance scores and recall frequency slow that decay. Frequently used memories resist forgetting; ignored ones get pruned automatically every 24 hours once strength drops below 0.05.
2. Four memory categories with different lifespans. Strategy memories (successful patterns) survive about 38 days without recall. Facts like preferences and identity last 24 days. Assumptions stick around 19 days. Failures, errors, and environment quirks fade in 11 days.
3. Hybrid retrieval combining vector, BM25, and graph search. Round one runs cosine similarity against stored memories. Round two does BFS graph traversal to surface related context that vocabulary search misses. A decayed memory survives if a graph neighbor is still strong, so related memories age together.
4. Three MCP tools called automatically. recall_memory fires at the start of every task, store_memory saves new learnings with an importance score, and update_memory replaces outdated facts. Once Claude has the CLAUDE.md instructions in place, the workflow runs without prompting.
5. Multi-agent support with isolated permissions. Multiple agents can share one YourMemory instance, each with its own private memories plus controlled access to shared context. API keys gate read/write permissions per agent.

How it compares to Zep Cloud

The headline number is 59% vs 28% recall on LoCoMo-10. Beyond the score, YourMemory runs locally with no Docker, no external services, and no managed database. Zep Cloud is a hosted service. For developers who want memory data on their own machine, that’s a meaningful tradeoff.

Availability and stack

YourMemory ships as a free open install via two commands. It supports Python 3.11 through 3.14. The default stack uses DuckDB for vector storage, NetworkX for the graph backend, sentence-transformers (all-mpnet-base-v2, 768 dims) for local embeddings, spaCy for NLP, and APScheduler for the automatic 24-hour decay job. PostgreSQL with pgvector and Neo4j are optional swaps for teams or larger datasets. The local database lives at ~/.yourmemory/memories.duckdb.

Practical use cases

• Coding agents that remember your stack, preferences, and past failures across sessions
• Teams sharing context between multiple agents without cross-contamination of private notes
• Long-running projects where session-by-session amnesia wastes setup time

Caveats

The tool requires a manual CLAUDE.md instruction file with your name and user ID for the recall-store-update workflow to fire automatically. VS Code-based clients like Cline don’t inherit shell PATH, so you’ll need the full executable path from yourmemory-path. The four decay categories are predefined, so anyone wanting custom curves would need to dig into the code.

For developers tired of re-explaining their stack to every new chat session, YourMemory tackles a real friction point with measurable results. Full methodology and per-sample breakdowns are available in the project’s BENCHMARKS.md at the original source.