Engram (DeepSeek)
https://arxiv.org/abs/2601.07372
DeepSeek + PKU. Introduces "conditional memory" alongside MoE's conditional compute. 27B-param N-gram lookup store offloads static knowledge from the transformer; O(1) lookup; prefetched from host RAM at near-zero runtime overhead.
Active · Verified 2026-10-23 · fresh as of 2026-10-23 · 33 cited facts, 1 with a third-party source · 1 recorded change · Suggest a correction
At a glance
- Type
- Conditional-memory N-gram lookup module vendor fix
- Tier
- T1
- Latest release
- no releases repo fix
- Pricing
- not applicable — research paper fix
- Funding
- not applicable — not commercial fix
- Freshness
- 2026-10-23 fix
- Layer
- agent-infra repo fix
- Licensing
- open-source repo fix
- Developer activity
- OpenRank 0.41 (12-mo avg 0.75) · 1 participants — 2026-09 third-party fix
- Maintainer health
- 2 issues in 90 days · 0 releases in 12 months · bus factor 1 — as of 2026-10-31 repo
Memory taxonomy
- storage
- parametric
- retrieval
- parametric-recall
- persistence
- parametric-permanent
- update
- read-only
- unit
- chunk
- governance
- opaque
- conflict
- n/a
When to use
Optimised for: not applicable - research paper
Anti-fit: not applicable - research paper
Pros & cons
Pros
DeepSeek-quality model with explicit memory architecture — open weights, strong on Chinese benchmarks; novel agentic memory design.
Cons
Documentation primarily in Chinese; smaller English-language adoption; eval coverage on Western benchmarks is limited.
Claims & capabilities
MMLU +3.4, CMMLU +4.0, BBH +5.0, ARC-Challenge +3.7. Multi-Query NIAH 84.2 → 97.0 vs iso-FLOPs MoE baseline.
Technical surface
- API surface
- searched not found
- Backend storage
- searched not found
- Deployment
- not applicable — research paper
- Embedding model
- searched not found
- Multi-tenancy
- searched not found
- MCP
- not applicable — research paper, no deployed product
- A2A
- not applicable — research paper, no deployed product
- OpenTelemetry
- not applicable — research paper, no deployed product
Compare Engram (DeepSeek) with…
Similar systems
Other recent method papers — theorized, no distinct product in the catalog, ranked by inbound references.
- Compressive Transformer T3
Maintains recent states in full resolution while compressing older memories with learned compression functions. DeepMind.
- MemGPT v2 / agent-tools T3
Already in catalog as the foundational MemGPT paper. Note: Letta is the productionised successor (cross-listed).
- Transformer-XL T3
Extends context through segment-level recurrence + caching of hidden states from prior segments. Foundational long-context architecture.
- Generative Agents T3
Park et al. — landmark agent-simulation paper. Reflection + memory stream + retrieval enable believable agent behavior.
- MemoryBank T3
Enhances LLMs with long-term memory. Early influential paper.
- Reflexion T3
Language agents with verbal reinforcement learning.
Sources (32 cited facts, 1 third-party)
Each fact links to the page it came from; the badge says who published it (how to read sources).
- What it is Conditional-memory N-gram lookup module vendor
- What it does & what's distinct DeepSeek + PKU. Introduces "conditional memory" alongside MoE's conditional compute. 27B… vendor
- Claims / benefits MMLU +3.4, CMMLU +4.0, BBH +5.0, ARC-Challenge +3.7. Multi-Query NIAH 84.2 → 97.0 vs iso… vendor
- Modalities text (LM training — O(1) N-gram conditional lookup) vendor
- Created 2026-01 vendor
- Latest release no releases repo
- License Apache-2.0 (code); Engram model use under a separate Model License repo
- GitHub deepseek-ai/Engram — ~4.6k stars, 4 commits, last commit not confirmed repo
- Mindshare Medium — several ML community blogs + LinkedIn posts about DeepSeek V4 architecture vendor
- Citations 28 cites (28/yr) paper
- Founders / pedigree DeepSeek (High-Flyer Capital spin-off); founder Liang Wenfeng; Chinese AI lab vendor
- Memory volume / scale Scaled to 27B parameters; MoE baseline comparison vendor
- Performance +3.7 poi ARC-Challenge +5.0 poi BBH +4.0 poi CMMLU +3.4 poi MMLU 84.2 NIAH (Multi-Query) vendor
- Reproducibility Too recent — Jan 2026 preprint; no independent reproductions vendor
- Code/weights release Code public + demo code repo
- Versioning none vendor
- Schema evolution schema locked vendor
- Pros DeepSeek-quality model with explicit memory architecture — open weights, strong on Chine… vendor
- Cons Documentation primarily in Chinese; smaller English-language adoption; eval coverage on … vendor
- Project & sources arxiv 2601.07372 GitHub deepseek-ai/Engram vendor
- obs-langsmith no vendor
- obs-opentelemetry no vendor
- obs-datadog no vendor
- obs-helicone no vendor
- obs-weave no vendor
- obs-langfuse no vendor
- obs-arize no vendor
- obs-custom Research paper — no production observability surface vendor
- commit-trajectory [{"ym":"2026-01","cum":4}] repo
- licensing-bucket open-source repo
- taxonomy-three-layer-rollup agent-infra repo
- developer-activity-opendigger {"activity":1.49,"as_of":"2026-09","openrank_12mo":0.75,"openrank_current":0.41,"partici… third-party
What changed
Changes to this record's lifecycle, licence, release, pricing and hosting facts, from the catalog's git history. Subscribe (Atom)