What it does SimpleMem is a long-term memory stack for LLM agents. It ingests text, images, audio, or video; compresses them into structured, semantically “lossless” memory units; and retrieves relevant context by meaning rather than keyword matching. The package auto-routes between text and multimodal backends based on your first API call.

The interesting bit The EvolveMem layer is the twist: it runs an LLM-driven Evaluate → Diagnose → Propose → Guard cycle on a dev set to discover retrieval hyperparameters—and even new retrieval dimensions—that weren’t in the original design. The README claims +25.7% relative improvement on LoCoMo and +18.9% on MemBench from this self-evolution, though independent verification is left as an exercise.

Key highlights

• Single pip install simplemem; mode="auto" picks text or omni backend from your first method call
• MCP server at mcp.simplemem.cloud plugs into Claude Desktop, Cursor, LM Studio, Cherry Studio, or any MCP client
• Multimodal support: text, image, audio, video with unified add_* / query interface
• Parallel processing knobs for batch ingestion and retrieval
• simplemem.optimize() exports a tuned Config JSON for production deployment

Caveats

• Benchmark claims (LoCoMo F1=0.613, “+47%”, “+51%”) are self-reported from the research paper; no third-party replication is cited in the README
• The “semantic lossless compression” framing is aspirational—what’s actually implemented is structured fact extraction with coreference resolution and absolute timestamps

Verdict Worth a look if you’re building persistent agents and tired of hand-tuning RAG pipelines. Skip it if you need battle-tested, externally validated retrieval—this is research code with a PyPI package and a cloud MCP endpoint.