Catalog Catalogs Memory Research & Benchmarks

Memory Research & Benchmarks

Papers, benchmarks and ideas behind agent memory, before they ship as products. For researchers and teams evaluating memory claims.

Method papers, benchmarks and evaluation suites, research systems and informal proposals, with citation counts, code release, reproducibility and the lineage and influence graph that connects them to shipping products.

264 systems · 4 sections · 6 acquired, archived, stale or shut down · data as of 2026-10-31

Browse the table → JSON · CSV

Decide

Most referenced

Lowest tier first, then by how many other systems in the catalog reference each one.

  1. Engram (DeepSeek) T1 Conditional-memory N-gram lookup module
  2. NVIDIA Nemotron 3 T1 Hybrid Mamba + sparse KV-cache, 1M-token context
  3. MemoryLLM (original) T2 Fixed-size in-weights memory pool with online self-update
  4. MetaGPT T2 Multi-agent collaborative framework
  5. Anthropic Circuit Tracing T2 Mechanistic interpretability of parametric storage
  6. Claude Plays Pokemon (Anthropic) T2 Anthropic agent demo — Claude 3.7/4 Sonnet plays Pokemon Red on Twitch
  7. IBM Meta-Tokens T2 Lossless context compression via learned placeholders
  8. ICLR 2026 MemAgents Workshop T2 Workshop venue — RL + cognitive psychology + LLM memory
  9. Mamba T2 Selective state-space model
  10. MemPalace T2 Verbatim chunks + spatial metaphor
  11. MineRL Diamond Challenge T2 Minecraft RL benchmark — obtain a diamond
  12. NAVER Provence + PISCO T2 RAG context pruning + soft compression

All 264 in the table →

Sections

Lifecycle

  • Active 233
  • Research code 25
  • Inactive 1
  • Archived 5

Tiers: T1 2 · T2 14 · T3 110 · T4 126 · T5 12

Recently changed

Everything that changed →

Research

Data and tools

  • JSON · CSV, flattened, one row per system, CC-BY-4.0
  • MCP tools: search_records, find_related, compare, system_history
  • CLI: landscape search · landscape related <id>
  • Data & API · Submit a system