Detailed Introduction
General Agentic Memory (GAM) is a general-purpose memory framework for agents that follows a Just-in-Time (JIT) memory optimization paradigm. At runtime, GAM performs deep research to retrieve and synthesize high-utility context from session data, using a dual-agent design — Memorizer and Researcher — to build structured memories and iteratively retrieve and refine context for downstream agents. The framework demonstrates strong performance on long-context benchmarks.
Main Features
- JIT memory optimization: retrieve and synthesize context at runtime to reduce up-front processing.
- Dual-agent architecture: Memorizer constructs memory events while Researcher performs iterative retrieval and summarization.
- Modular design: pluggable retrievers, embedding generators and model backends for flexible integration.
- Reproducible evaluation: includes an evaluation suite to reproduce LoCoMo, HotpotQA, RULER and NarrativeQA results.
Use Cases
- Agent memory enhancement: provide long-term, context-aware memory for conversational and autonomous agents.
- Research & benchmarking: reproduce paper experiments and evaluate memory strategies.
- Hybrid deployment: run locally or in-cloud for research and production scenarios.
Technical Features
- Retrieval strategies: support dense vector retrieval and BM25 hybrid approaches.
- Cross-model compatibility: works with cloud LLMs and local runtimes (e.g., OpenAI, vLLM).
- Evaluation-driven: unified CLI and scripts for running benchmarks and reproducing results.
- Open-source license: MIT license for research reuse and engineering extension.