SimpleMem: Efficient Lifelong Memory for LLM Agents
Store, compress, and retrieve long-term memory with semantic lossless compression — text and multimodal — so agents recall more while spending fewer tokens.
Evidence shows: README requires users to configure API keys but does not explain key storage or least privilege; MCP server supports multi-tenant isolation and token auth but lacks user confirmation; data flow transparency is limited, not specifying how data is stored, transmitted, or deleted; sensitive data handling mentions ENCRYPTION_KEY but lacks detail; dependency security pins versions but no vulnerability scanning; external effects: cloud-hosted MCP but no data residency or privacy policy; rollback: EvolveMem mentions automatic rollback but not detailed; source attribution: paper and citations present but unverified. Deductions: lack of user confirmation, insufficient data flow transparency, missing sensitive data handling details, unverified dependency security, unspecified external effects, unclear rollback, unverified source attribution.
Evidence shows: README and tests are mostly consistent, but minor inconsistencies like 'semantic lossless compression' not tested; dependency availability: requirements pinned but not verified; failure messages: README gives one tip but no comprehensive error handling docs. Deductions: slight inconsistency, unverified dependency availability, incomplete failure messages.
Evidence shows: README clearly defines target audience (LLM agent developers) and scenarios (text and multimodal memory), offers multiple integration methods (Python, MCP, Docker); capability boundaries: lists supported and planned features but not limitations; trigger precision: auto mode selects backend based on first call but not detailed; environment fit: supports various OpenAI-compatible endpoints but not all configurations. Deductions: trigger precision not detailed, capability boundaries not fully explicit.
Evidence shows: README well-structured with quick start, installation, MCP server, reproduction sections; install notes detailed including pip and Docker; naming stability: unified 'simplemem' package but no version history; examples and FAQ: multiple code examples but no FAQ; known limitations: mentions MCP lacks multimodal and EvolveMem but not comprehensive; license: MIT clear; versioning: news updates but no formal changelog; maintenance responsibility: no maintainer or contribution guidelines. Deductions: naming stability unverified, known limitations incomplete, versioning informal, maintenance responsibility unclear.
Evidence shows: output usability: code examples with expected outputs but not verified; marginal value: claims SOTA on benchmarks but no reproducible details; cost benefit: claims token reduction but no specific data. Deductions: outputs unverified, marginal value insufficiently supported, cost benefit data lacking.
Evidence shows: README cites arXiv paper and benchmark results but no experimental details; cross-source corroboration: only README and tests, no independent verification; fact-inference separation: claims and inferences mixed. Deductions: claims lack traceability, insufficient cross-source corroboration, facts and inferences not separated.
- API key handling is unclear; review code to ensure secure storage.
- Cloud-hosted MCP server may raise data privacy concerns; verify data residency and privacy policy.
- Dependencies are pinned but no vulnerability scanning; regularly check dependency security.
- Claimed benchmark performance lacks reproducible details; treat with caution.
What does this agent do, and when should you use it?
SimpleMem is a unified memory stack for LLM agents, built on storing semantically lossless memory at high information density to boost recall while cutting inference token usage by roughly 30x. It bundles three components: SimpleMem (text core), Omni-SimpleMem (multimodal memory for text, image, audio, video), and EvolveMem (self-evolving retrieval). Deployable via a Python package (simplemem) or an MCP server (with Docker self-hosting), it integrates with clients like Claude Desktop and Cursor. It works with any OpenAI-compatible API and reports state-of-the-art results on LoCoMo, MemBench, and Mem-Gallery. The repo is MIT-licensed under aiming-lab.
SimpleMem implements a three-stage pipeline: store (converts interactions into compact, structured memory units), index (organizes memories with semantic embeddings and metadata), and retrieve (fetches relevant memories by meaning). The Python API offers SimpleMem(), add_dialogue(), add_text(), add_image(), add_audio(), add_video(), ask(), and query(). It supports parallel processing (enable_parallel_processing) and self-optimization via simplemem.optimize(). The MCP server exposes text memory over streamable HTTP (MCP 2025-03-26 protocol) with hybrid retrieval and multi-tenant isolation, and is available at mcp.simplemem.cloud.
- Enhancing chatbots or virtual assistants that need cross-session context.
- Adding long-term memory to multimodal analysis applications that reference past images or audio.
- Processing large dialogue datasets more efficiently by reducing token overhead for context.
- Integrating memory capabilities into existing AI workflows via MCP or Python API.
- Improving retrieval accuracy with self-evolving configuration tuning on custom dev sets.
What are this agent's strengths and limitations?
- Reports SOTA on LoCoMo (F1=0.613, +47%) and reduces inference tokens by ~30x.
- True multimodal support (text, image, audio, video) via a unified Python API.
- Self-evolving retrieval (EvolveMem) with +25.7% relative gain on LoCoMo.
- Flexible deployment: Python package, MCP cloud service, and self-hosted Docker.
- Requires a paid OpenAI-compatible API key; must be configured before any use.
- Setup complexity: requires Python 3.10+ and configuration knowledge.
- Multimodal and MCP are separate offerings; multimodal via MCP is not yet available.
- Benchmark claims (e.g., F1) are self-reported; independent validation data not provided in the repo.
How do you install or deploy this agent?
Clone the repo: git clone https://github.com/aiming-lab/SimpleMem.git and enter it. Install dependencies with pip install -r requirements.txt or pip install -e . (default). Optional extras: pip install -e .[server] for MCP/HTTP, pip install -e .[benchmark] for reproduction. Requires Python 3.10+ and an OpenAI-compatible API key. Copy config.py.example to config.py and set OPENAI_API_KEY (and optionally OPENAI_BASE_URL). For Docker, run docker compose up -d.
How do you use this agent?
For text memory, instantiate SimpleMem(), add dialogues with add_dialogue(), then finalize() and query with ask(). For multimodal, use add_text(), add_image(), etc. To optimize retrieval, call simplemem.optimize(mem, dev_questions, max_rounds=3) and save the config. For MCP, configure mcpServers with the URL and a Bearer token. See the repo README and docs for detailed examples.