MSA (Memory Sparse Attention) — An Open-Source Answer to AI’s Memory Bottleneck

The Core Problem

Where AI memory stands today:

  • The strongest models have an effective context of roughly 1M tokens
  • The information a human can store in a lifetime is on the order of 200–300 million tokens
  • That’s a gap of two orders of magnitude

Both of the industry’s well-worn paths have hit a wall:

  1. Stretching the context window — compute cost grows quadratically; that road ends here
  2. Bolting on RAG — retrieval and generation stay disconnected, and accuracy has a ceiling

What MSA Is

MSA (Memory Sparse Attention) comes from the EverMind team (under Shanda). It embeds memory directly into the attention mechanism itself — no longer context, no external retrieval.

In one sentence: traditional RAG gives a model an external hard drive; MSA gives it a native memory chip.

  • Finding and using are no longer two separate steps — they’re fused into a single neural network and done end to end
  • The model learns on its own what to remember, how to find it, and how to use it
  • Plug and play: you only swap out the Self-Attention layer of a standard Transformer

Key Technical Details

  1. Compression mechanism: brings 100 million tokens of storage down to an acceptable footprint
  2. Tiered storage: routing indexes on the GPU, content details on the CPU — total capacity depends on memory, not VRAM
  3. Sparse routing: complexity drops from O(L²) to O(L)
  4. Positional encoding: each document is numbered independently, so training at 64K extrapolates to 100M

Performance

Built on Qwen3-4B, with 159B tokens of continued pre-training:

Result Data
Memory span from ~10K tokens → 100 million tokens (nearly 4 orders of magnitude)
Quality degradation answer quality drops only <9%
Standard QA benchmarks a 4-billion-parameter model beats traditional RAG approaches by 16%
vs. top-tier retriever + 235-billion-parameter LLM still wins on multiple tests, with a 60× parameter gap

Hardware Bar

It runs directly on a machine with two A800 GPUs — no cluster needed. Small and mid-sized teams, even individual developers, can have long-term memory at the hundred-million-token scale.

Team Background

  • EverMind (under Shanda)
  • Previously built Omne, the multi-agent framework that reached SOTA on the GAIA leaderboard, and EverOS, an open-source memory platform
  • More than nine months from project kickoff to completed paper

Key insight: a model needs different information when it’s “finding material” versus “writing the answer” — finding calls for macro-level judgment, writing calls for micro-level detail. Split them apart and handle each with a dedicated module, and performance changes qualitatively.

Application Prospects

  • Genuinely personalized AI assistants: remembering dietary preferences, project progress, family members’ personalities
  • AI education: truly personalized, teaching to the individual
  • Medical assistants: tracking a complete medical history
  • Enterprise knowledge bases: remembering a decade of accumulated project knowledge
  • Memory as a service: the memory layer as an independent, pluggable module, so memory assets aren’t locked to a single model