Engram
An internal memory mechanism introduced in DeepSeek-V4.1-Flash that allocates a portion of the model's weight capacity—197B parameters—as an n-gram memory table. High-frequency language patterns are handled via lookup rather than forward computation, reducing inference cost.
Core Facts
Timeline (last 90 days)
Betelgeuse's 110 GB Engram table is placed entirely on the NVMe SSD, is activated only in layers 1 and 14, touches only 48 rows (about 192 KB) per token, and the active rows are staged into a 2MB VRAM buffer at t=0.
Compressing a 763B-class MoE model to 72 GB of transformer weights means consumer-grade or single-machine multi-GPU environments now have the possibility of running an ultra-large model, but the 110.6 GB Engram file remains a non-negligible extra overhead.
The shared Engram file is 110.6 GB, exactly the same as the unpruned version, and the vision projector is also the same as the unpruned version, so it only needs to be downloaded once.
The Engram n-gram memory module of DeepSeek-V4.1-Flash occupies 196 billion parameters.
DeepSeek 1017's expert weights are already FP4 (4 bits per parameter), so conventional 4-bit quantization is of little help; the Engram memory table can be placed on SSD (reading only about 12KB per token), but the expert weights must reside in high-speed memory, because the router selects 6 of 384 experts per layer and makes a different selection for each token.
Engram is a graph-database-based shared memory system built by developers at FalkorDB's Graph Hacks hackathon, providing shared memory for all AI tools.
Engram represents a new memory superseding an old one via a SUPERSEDES edge; old memories are marked superseded but not deleted, and recall filters to active status by default.
Engram's provenance path is Tool -[:RAN]-> Session -[:CAPTURED]-> Memory -[:SUPERSEDES]-> Memory, achieving provenance by traversing edges without needing synchronized JSON fields.
Engram performs one-hop traversal via RELATED_TO entity edges, enabling recall of memories that don't directly mention the query terms, because entity co-occurrence strengthens the edges.
Engram automatically captures memories via Claude Code's session hooks, with no need for users to explicitly say "remember this" — decisions land in the graph at session end.
16 more timeline events
All Facts (20)
DeepSeek-V4.1-Flash's KV cache is compressed to 890 bytes per token, only one-quarter that of V4 Flash.
70%UnverifiedBetelgeuse's 110 GB Engram table is placed entirely on the NVMe SSD, is activated only in layers 1 and 14, touches only 48 rows (about 192 KB) per token, and the active rows are staged into a 2MB VRAM buffer at t=0.
50%UnverifiedCompressing a 763B-class MoE model to 72 GB of transformer weights means consumer-grade or single-machine multi-GPU environments now have the possibility of running an ultra-large model, but the 110.6 GB Engram file remains a non-negligible extra overhead.
50%UnverifiedThe shared Engram file is 110.6 GB, exactly the same as the unpruned version, and the vision projector is also the same as the unpruned version, so it only needs to be downloaded once.
50%UnverifiedThe Engram n-gram memory module of DeepSeek-V4.1-Flash occupies 196 billion parameters.
50%UnverifiedDeepSeek 1017's expert weights are already FP4 (4 bits per parameter), so conventional 4-bit quantization is of little help; the Engram memory table can be placed on SSD (reading only about 12KB per token), but the expert weights must reside in high-speed memory, because the router selects 6 of 384 experts per layer and makes a different selection for each token.
50%UnverifiedEngram is a graph-database-based shared memory system built by developers at FalkorDB's Graph Hacks hackathon, providing shared memory for all AI tools.
50%UnverifiedEngram represents a new memory superseding an old one via a SUPERSEDES edge; old memories are marked superseded but not deleted, and recall filters to active status by default.
50%UnverifiedEngram's provenance path is Tool -[:RAN]-> Session -[:CAPTURED]-> Memory -[:SUPERSEDES]-> Memory, achieving provenance by traversing edges without needing synchronized JSON fields.
50%UnverifiedEngram performs one-hop traversal via RELATED_TO entity edges, enabling recall of memories that don't directly mention the query terms, because entity co-occurrence strengthens the edges.
50%UnverifiedEngram automatically captures memories via Claude Code's session hooks, with no need for users to explicitly say "remember this" — decisions land in the graph at session end.
50%UnverifiedCursor and Gemini CLI read the same Engram graph via MCP, but require manually adding the engram_recall prompt because they lack a hook system.
50%UnverifiedEngram's local web UI draws the real graph, with nodes colored by status and SUPERSEDES edges highlighted, showing real data rather than a conceptual diagram.
50%UnverifiedIn the Engram approach, MCP reads rely on the prompt's self-discipline, and automatic capture is available only in tools that support hooks — these are the current approach's limitations.
50%UnverifiedEngram is a hash-based n-gram memory stored in host memory and attached to 2 of the 72 layers; mHC replaces the traditional single residual stream with 4 residual streams.
50%UnverifiedEngram is a hardware device that combines a sampler and a groovebox, using AI to mangle input audio and generate entirely new sounds.
50%UnverifiedEngram combines a sampler with a groovebox, aiming to cover the complete chain from sound capture and AI processing to rhythm arrangement.
50%UnverifiedEngram positions AI as a "new type of instrument" that expands human capability rather than an "automatic composition machine" that replaces humans.
50%UnverifiedEngram explicitly states that it is not "Suno in a box" and is not a one-click song-generation device.
50%UnverifiedEngram presents AI audio processing in a groovebox form factor, emphasizing that performers make decisions in the moment rather than repeatedly tweaking parameters in a software interface.
50%