GAM: General Agentic Memory

BAAI's agent-memory framework, from the VectorSpaceLab team led by Zheng Liu, and the first in the line that JAM continues. Most memory systems for agents (A-MEM, Mem0, MemoryOS, LightMem) compress an agent's history into notes or summaries ahead of time, which GAM's authors liken to ahead-of-time compilation and blame for losing details that later requests need. GAM works just in time instead. An LLM Memorizer writes a short memo for each session into a lightweight memory and stores the full session, prefixed with a context header, as a page in a page-store. For each request, an LLM Researcher plans searches over the pages (BGE-M3 embeddings, BM25 and direct page lookup), integrates what it finds, and checks whether the evidence suffices, repeating up to three times before returning a compact context to the client agent. The paper formulates end-to-end RL for both roles, but its experiments only prompt the models; JAM (September 2026) is the version that trains the Researcher.

With GPT-4o-mini as the backbone for GAM and every baseline, GAM scores 57.8, 42.3, 59.5 and 33.3 F1 on LoCoMo's single-hop, multi-hop, temporal and open-domain questions, against best baselines of 52.5 (RAG), 38.7, 48.9 and 28.6 (all Mem0). On HotpotQA it scores 63.2, 64.6 and 59.8 F1 at 56K, 224K and 448K tokens, against 52–57 for plain long-context prompting and RAG and 23–41 for the memory systems. On RULER's 128K multi-hop tracing it reaches 93.2%, where RAG and A-MEM score zero, and on NarrativeQA 36.9 F1 against 31.3. Results with Qwen2.5-14B-Instruct follow the same pattern. The two roles need different model sizes: a 0.5B Memorizer costs little (48.8 against 53.2 average F1 with 14B), but a 0.5B Researcher collapses to 9.1. The research step adds 12–18 s per HotpotQA request, against under a second for the memory baselines, though GAM's total time including memory construction (69 s at 56K, 576 s at 448K) stays below A-MEM's and MemoryOS's. The repository has since grown into a general agent file-system framework, with a Python SDK, CLI, REST API, web interface, video memory and long-horizon trajectory compression, and keeps the paper's code in a research directory; it has about 860 GitHub stars. The package declares an MIT license, though the repository has no LICENSE file.