TIMAN-group/PlugMem

What it solves

PlugMem addresses the problem of context overhead and inefficient memory retrieval in LLM agents. Instead of relying on raw interaction histories, it provides a system to organize experience into compact, reusable knowledge units that agents can recall based on what is actually relevant to decision-making.

How it works

PlugMem operates as a task-agnostic memory backbone that can be integrated into existing agent pipelines. It uses a graph-based structure to maintain hierarchical knowledge units across three distinct memory types:

• Semantic Memory: Stores facts and concepts, such as user preferences and factual information.
• Procedural Memory: Stores workflows and step-by-step processes (how-to knowledge).
• Episodic Memory: Stores long interaction sequences on disk, referenced by ID.

The system leverages LLMs for knowledge extraction, retrieval, and reasoning, and supports the evolution and updating of the memory graph over time.

Who it’s for

Developers building LLM agents for dialogue, knowledge-intensive QA, and web automation, as well as those using AI coding agents like Claude Code or OpenClaw.

Highlights

• Plug-and-Play Integration: Can be added to agents with minimal code changes.
• Memory Inspector: A built-in web UI for visualizing the graph topology, browsing memories, and debugging retrieval.
• Coding Agent Support: Includes a plugin for Claude Code that allows the agent to learn project conventions and debugging recipes across sessions.
• State-of-the-Art Performance: Achieves SOTA results on LongMemEval and HotpotQA benchmarks.