Your personal memory, across sessions, agents, and devices.

DeepSeek V4 Engram Memory: What Actually Shipped

memU Team memU Team

When this post first ran in February 2026, it said DeepSeek V4 would ship with Engram. Engram is a "conditional memory" module from a January DeepSeek paper. V4 shipped on April 24, 2026, and its public model card does not mention Engram. V4 is a pair of open-weight models built for a 1M-token context. Engram is a research idea about where a model stores fixed knowledge inside its weights. Neither one remembers what your coding agent learned about your project last week, and the last section covers that gap.

What are DeepSeek V4 and Engram, in one sentence each?

DeepSeek V4 is a family of open-weight Mixture-of-Experts models with a 1M-token context window. Engram is a DeepSeek research module that adds a fast lookup table for static patterns next to a model's computation.

DeepSeek announced the V4 preview on April 24, 2026. V4-Pro has 1.6 trillion total parameters with 49 billion active. V4-Flash has 284 billion total with 13 billion active. Both support Thinking and Non-Thinking modes. DeepSeek says 1M-token context is now the default across its official services.

Engram comes from the paper Conditional Memory via Scalable Lookup, first submitted on January 12, 2026. Its abstract describes Engram as a module that modernizes classic N-gram embeddings for constant-time lookup.

How do V4's long context and Engram's memory differ?

DeepSeek V4 long context Engram conditional memory Agent memory (outside the model)
What it is A 1M-token window the model reads per request A lookup table of N-gram embeddings inside the model Notes and skills kept between sessions
Where it lives The request, plus a KV cache on the server Model weights, which can sit in host memory Files or a store you control
Who fills it You or your agent, on every request Training The agent, during and after sessions
Survives a new session? No. Each request starts from what you send Yes, but only general patterns learned in training Yes
Shipped in V4? Yes Not mentioned in the V4 model card Not a model feature

Sources: DeepSeek's V4 announcement and V4-Pro model card. Engram from the paper.

How does DeepSeek V4 handle a 1M-token context?

The V4-Pro model card describes a hybrid attention design. It combines Compressed Sparse Attention with Heavily Compressed Attention. It also uses Manifold-Constrained Hyper-Connections for residual connections and the Muon optimizer for training. Pre-training used more than 32 trillion tokens.

DeepSeek's headline efficiency claim is about cost at full length. At 1M tokens, the card says V4-Pro needs 27% of the single-token inference FLOPs of DeepSeek-V3.2 and 10% of its KV cache. The card also reports 83.5 on MRCR 1M, which tests long-context retrieval. In its Pro-Max mode it reports 80.6 on SWE-bench Verified. We found no independent reproduction of these numbers.

The model card and repository carry the MIT license. On the API, V4 takes the OpenAI ChatCompletions format and the Anthropic format at the same base URL. DeepSeek says V4 works with Claude Code and OpenClaw, among other agents. The older deepseek-chat and deepseek-reasoner names were retired after July 24, 2026.

A V4.1-Flash announcement on September 10, 2026 describes a 552-billion-parameter MoE with native visual understanding. In October the API's deepseek-flash name moved to DeepSeek-V4.1-Flash-1017, with weights on Hugging Face.

How does Engram work?

The paper's abstract starts from one observation. Transformers have no native lookup step, so they spend computation rebuilding facts they could look up. Mixture-of-Experts scales capacity through conditional computation. Engram adds a second axis, conditional memory, as a table the model can index directly.

The authors frame the split between MoE experts and Engram memory as a "Sparsity Allocation" problem. They report a U-shaped scaling law for it and scale Engram to 27 billion parameters. Against a MoE baseline with the same parameters and FLOPs, they report gains such as +5.0 on BBH and +3.0 on HumanEval. Multi-Query Needle-in-a-Haystack rose from 84.2 to 97.0. The largest gains came in reasoning and code. The authors credit Engram with freeing early layers from static pattern work.

Because Engram's addressing is deterministic, the abstract says the table can be prefetched from host memory at runtime with negligible overhead. A large knowledge table could then live in cheaper CPU memory. The Engram repository ships a demonstration only. Its README says the code illustrates the data flow and mocks the other components.

Does DeepSeek V4 use Engram?

The official V4 sources we read do not say so. The V4 announcement describes token-wise compression plus DeepSeek Sparse Attention. Neither it nor the V4-Pro model card, which lists hybrid attention and the Muon optimizer, mentions Engram. The Engram README does not mention V4 either.

Early-2026 claims of "V4 with Engram memory" were predictions the release did not confirm. DeepSeek revised the Engram paper in July 2026, so the work is still active. It may yet appear in a later model.

What should you read in the fine print?

Every benchmark above is self-reported. The V4 numbers come from DeepSeek's model card. The card calls Pro-Max the best open-source model, which is DeepSeek's claim. The Engram gains are measured against the paper's own baseline at 27 billion parameters, which is far smaller than V4-Pro.

Long context also has a cost on the calling side. DeepSeek's context caching guide says caching is on for all users and matches on exact prefixes. Unused caches are cleared within a few hours to a few days. A cache saves money only while you keep resending the same prefix.

The bottom line

Use DeepSeek V4 for what it shipped: MIT-licensed open weights and a 1M-token window that DeepSeek says costs less than V3.2's. Do not plan around Engram until a released model card names it. Neither one keeps what your agent learned between sessions, so project memory still has to live outside the model.

Does a 1M-token model remember your project between sessions?

No. The 1M-token window holds only what you send in that request, and nothing carries over to the next one. Engram, if it ships, holds general patterns learned in training. Neither one knows that your repository pins a library version or that the last migration failed on a foreign key. Your agent has to send that context again every session, or rediscover it. Since V4 runs inside Claude Code and OpenClaw, one model can sit behind several hosts. Each host keeps its own notes.

memU keeps that project knowledge outside the model, as a shared wiki of readable Markdown. A scheduled background task slices new session logs from each host into jobs. The agent decides what is worth keeping and writes memory or skill Markdown. memU makes no LLM or chat calls itself. A standing instruction in CLAUDE.md or AGENTS.md tells the agent to retrieve before answering. Because memU binds to the host, the notes stay put when you swap the model behind Claude Code from Claude to DeepSeek. All hosts share one backend, so a note written from Claude Code can be retrieved in Codex.

To try the hosted version, which is free and syncs across devices, get an API key from memu.so and send your agent one message:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is YOUR_MEMU_API_KEY.

The self-hosted setup and the support matrix are in the memU GitHub repository.

FAQ

Does DeepSeek V4 have Engram memory?

Not according to DeepSeek's published V4 sources. The V4-Pro model card describes hybrid compressed attention and never mentions Engram.

What is DeepSeek Engram?

Engram is a module from a January 2026 DeepSeek paper. It adds a lookup table of N-gram embeddings to a model, so fixed patterns are retrieved instead of recomputed.

What is DeepSeek V4's context length?

Both V4-Pro and V4-Flash support a 1M-token context, which DeepSeek made the default across its official services.

Is DeepSeek V4 open source?

The weights are on Hugging Face, and the V4-Pro model card states the MIT license for the repository and the weights.

How many parameters does DeepSeek V4 have?

V4-Pro has 1.6 trillion total parameters with 49 billion active per token. V4-Flash has 284 billion total with 13 billion active.