IAAR-Shanghai/Awesome-AI-Memory

What it is A community-curated repository that collects and organizes research papers, open-source projects, and practical knowledge about giving large language models (LLMs) persistent memory. The project exists because LLMs today rely on a limited "context window" — the amount of text they can see at once — which fundamentally makes them short-term-memory-only, unable to sustain long conversations, personalization, or multi-stage tasks on their own.

The repository aims to be a continuously evolving knowledge base for researchers and practitioners building AI systems that can remember over the long term and keep improving. (README: "仓库目标" — goals section)

What's inside

• A paper list of over 1,270 research papers (README badge: "Papers 1271"), categorized into surveys, method/framework papers, dataset/benchmark papers, and model/system papers.
• An open-source projects section with 111 projects (README badge: "Open Source Projects 111").
• A test/benchmarks section (📊 测试基准), multimedia resources (🎥 多媒体资源), and an "Adam framework" section (🧠 Adam 框架) — the README lists section headers but doesn't describe their contents further.
• A core-concepts cheat sheet defining memory types (short-term, long-term, episodic, parametric, etc.) and memory operations (write, retrieve, update, delete, compress).
• A change log showing regular updates — dozens of papers added per update, since the repository's initial release in December 2025.
• Contributing is encouraged (README badge: "PRs Welcome") via a documented screening process in SCREENING.md.

Scope — what gets in and what gets excluded (README "项目范围" section)

• Focus is on agent memory and internal memory mechanisms of language models — memory must be the core research question of a paper, studied on an LLM or multimodal LLM.
• Includes: long-term, episodic, semantic, shared, and experience-derived procedural memory; storage, retention, addressing, recall, and forgetting in LLMs; evaluations, security research, and surveys directly on these mechanisms.
• Excludes: pure vision or low-level robot state, hardware/KV-cache efficiency, generic RAG/long-context/tool routing, and adjacent work that only claims to transfer to LLMs.
• For papers where evidence is insufficient, the rule is to mark them for later review rather than reject or accept hastily — and to state precisely what was verified (e.g., not presenting abstract-level checks as full-paper checks).

How it's organized (as reflected in the README's structure and content) Papers are listed by publication date, grouped into categories (surveys; method/framework; dataset/benchmark; model/system). Each entry shows date, title, summary bullets, and an arXiv link. The taxonomy section in the README (currently commented out) would classify memory along dimensions like storage location (parametric vs. external), temporal scope (short vs. long-term), content type (episodic vs. semantic vs. procedural), and system architecture (RAG, vector retrieval, symbolic-neural hybrid).

Why it matters The README frames memory systems as solving LLMs' fundamental context-window limitation by adding external, persistent, controllable memory structures — allowing models to store, retrieve, compress, and manage historical information across sessions, achieving consistency and continuous reasoning beyond a single prompt.