Memory Research & Benchmarks

Recent method papers — theorized, no distinct product

Conditional-memory N-gram lookup module ↗ 2

Code public + demo code ↗ ●

Too recent — Jan 2026 preprint; no independent reproductions ↗ 2

Apache-2.0 (code); Engram model use under a separate Model License ↗ ●

2026-10-23

Recent method papers — theorized, no distinct product

Hybrid Mamba + sparse KV-cache, 1M-token context ↗ 2

Code public + weights public (model weights on HuggingFace; training recipes at NVIDIA-NeMo/Nemotron) ↗ 2

Too recent (weights public; training recipes released at github.com/NVIDIA-NeMo/Nemotron) ↗ ●

2026-10-28

Recent method papers — theorized, no distinct product

Mechanistic interpretability of parametric storage ↗ 2

Code public + weights closed (open-sourced circuit-tracing library; Neuronpedia frontend; no model weights released) ↗ 2

Original team only (Anthropic; open-source tools released June 2025; community replication pending) ↗ 2

2026-10-28

Memory benchmarks & evaluation

Anthropic agent demo — Claude 3.7/4 Sonnet plays Pokemon Red on Twitch ↗ 2

Source not released — Twitch livestream + Anthropic blog only ↗ 2

Twitch live-stream playthrough provides timestamped reproducibility evidence ↗ 2

2026-10-28

Recent method papers — theorized, no distinct product

Lossless context compression via learned placeholders ↗ 2

Paper only — no public code repo found ↗ 2

Too recent (ICML 2025 paper; no independent reproductions found) ↗ 2

2026-10-28

Memory benchmarks & evaluation

Workshop venue — RL + cognitive psychology + LLM memory ↗ 2

2026-10-23

Recent method papers — theorized, no distinct product

Selective state-space model ↗ 2

searched not found ↗ 2

searched not found ↗ 2

searched not found ↗ 2

2026-10-28

Recent method papers — theorized, no distinct product

Fixed-size in-weights memory pool with online self-update ↗ 2

Active — last commit 2025-07 · 314★ ↗ ●

Independently reproduced (ICML 2024 published; community replications confirmed) ↗ ●

2026-10-28

Research / specialised systems

Verbatim chunks + spatial metaphor ↗ 2

searched not found ↗ ●

searched not found ↗ ●

MIT ↗ ●

2026-10-28

Recent method papers — theorized, no distinct product

Multi-agent collaborative framework ↗ 2

Active — last commit 2026-01 · 67,918★ ↗ ●

Independently reproduced (ICLR 2024 oral — top 1.2%) ↗ 2

MIT ↗ ●

2026-10-23

Memory benchmarks & evaluation

Minecraft RL benchmark — obtain a diamond ↗ 2

Code public + Minecraft proprietary ↗ 2

Independently reproduced — DreamerV3 (DeepMind 2023), VPT (OpenAI 2022) ↗ 2

MIT ↗ 2

2026-05-14

Recent method papers — theorized, no distinct product

RAG context pruning + soft compression ↗ 2

searched not found ↗ 2

Too recent ↗ 2

2026-10-28

Memory benchmarks & evaluation

Roguelike RL benchmark — NetHack as agent environment ↗ ●

Code public + game-engine source open ↗ ●

Independently reproduced — community NetHack Challenge 2021 ↗ ●

NetHack General Public License ↗ ●

2026-05-14

Recent method papers — theorized, no distinct product

Per-neuron history + neural synchronization as latent ↗ 2

Code public + weights public (SakanaAI/continuous-thought-machines GitHub; model checkpoints released) ↗ ●

Too recent (code public at SakanaAI/continuous-thought-machines with checkpoints) ↗ ●

Apache-2.0 ↗ ●

2026-10-23

Recent method papers — theorized, no distinct product

KL-distillation experience internalization ↗ 2

Active — last commit 2025-05 · 15★ ↗ ●

Too recent ↗ 2

2026-10-28

Recent method papers — theorized, no distinct product

Low-rank CPU-offloaded KV cache + sparse reconstruction ↗ 2

Active — last commit 2025-05 · 297★ ↗ ●

Independently reproduced (ICML 2025 Spotlight acceptance implies community review) ↗ ●

Apache-2.0 ↗ ●

2026-10-28

Research / specialised systems

Atomic-note / Zettelkasten-style ↗ ●

Code public + weights closed ↗ ●

Too recent ↗ 2

MIT ↗ ●

2026-10-28

Recent method papers — theorized, no distinct product

Paper ↗ 2

searched not found ↗ 2

searched not found ↗ 2

Apache-2.0 (official code) ↗ 2

2026-10-23

Recent method papers — theorized, no distinct product

Workflow as memory ↗ 2

Code public + weights closed — https://github.com/zorazrw/agent-workflow-memory · 426* / 50 forks · last commit 2025-12 · Apache-2.0 ↗ ●

Independently reproduced: 50 forks of https://github.com/zorazrw/agent-workflow-memory; downstream papers cite results ↗ ●

2026-10-23

Memory benchmarks & evaluation

Benchmark ↗ ●

searched not found ↗ ●

searched not found ↗ ●

Apache-2.0 ↗ ●

2026-10-29

Recent method papers — theorized, no distinct product

Test-time plan-template memory ↗ 2

depth-floor-reached: arXiv 2506.14852 preprint; no official code repo discovered via WebSearch + arXiv abstract scan + GitHub API ↗ 2

No reproductions found — published 2025-06-17 (11 mo ago); no public code ↗ 2

2026-10-23

Recent method papers — theorized, no distinct product

Hopfield + low-rank explicit memory ↗ 2

depth-floor-reached: arXiv 2309.12862 preprint; no official code repo discovered via WebSearch + arXiv abstract scan + GitHub API ↗ 2

depth-floor-reached: no reproductions found via Semantic Scholar / WebSearch / arXiv abstract scan ↗ 2

2026-10-28

Recent method papers — theorized, no distinct product

Null-space knowledge editing ↗ 2

Code public + weights closed — https://github.com/jianghoucheng/AlphaEdit · 439* / 46 forks · last commit 2025-10 · MIT ↗ ●

Highly reproduced: ICLR 2025 Outstanding Paper; downstream method AlphaEdit+ (ICLR 2026) builds directly on it; 191 cites (mostly downstream); 46 forks of official repo ↗ ●

MIT ↗ ●

2026-10-28

Memory benchmarks & evaluation

Mobile-OS agent benchmark — 116 real Android app tasks ↗ ●

Code public ↗ ●

Independently reproduced — M3A, AppAgent-v2 evaluations ↗ ●

Apache-2.0 ↗ ●

2026-10-28

Recent method papers — theorized, no distinct product

CLS dual-layer episodic + semantic KG ↗ 2

Code public + weights closed — https://github.com/AIRI-Institute/AriGraph · stale (last commit 2024-09) · 166* / 28 forks · MIT ↗ ●

Original team only — 28 forks of https://github.com/AIRI-Institute/AriGraph (mostly fork-only; no public re-implementations identified via WebSearch) ↗ ●

MIT ↗ ●

2026-10-28

Memory benchmarks & evaluation

Sample-efficient RL benchmark — 100k env steps on Atari ↗ 2

Code public (ALE / atari-py) ↗ 2

Independently reproduced — multiple SOTA papers reference the protocol ↗ 2

GPL-2.0 (ALE) ↗ 2

2026-10-28

Memory benchmarks & evaluation

Distributed multi-fact reasoning across long docs ↗ 2

Code public + weights closed ↗ ●

Independently reproduced ↗ 2

Apache-2.0 ↗ ●

2026-10-28

Research / specialised systems

Hierarchical: short / mid / long-term ↗ ●

Code public + weights closed ↗ ●

Too recent ↗ 2

Apache-2.0 ↗ ●

2026-10-28

Memory benchmarks & evaluation

Benchmark ↗ 2

searched not found ↗ 2

searched not found ↗ 2

MIT code; CC BY-SA 4.0 data ↗ 2

2026-10-23

Memory benchmarks & evaluation

Benchmark ↗ 2

searched not found ↗ 2

searched not found ↗ 2

Apache-2.0 ↗ 2

2026-10-29

Memory benchmarks & evaluation

Benchmark ↗ 2

searched not found ↗ 2

searched not found ↗ 2

MIT ↗ 2

2026-10-29

Recent method papers — theorized, no distinct product

Thought-augmented reasoning ↗ 2

Code public + weights closed — https://github.com/YangLing0818/buffer-of-thought-llm · 676* / 64 forks · last commit 2025-06 · MIT ↗ ●

Original team + downstream attention: 64 forks of https://github.com/YangLing0818/buffer-of-thought-llm; published 2024-06-06 (23 mo ago); no full re-implementations identified ↗ ●

2026-10-23