Hybrid recall
BM25, dense vectors, ColBERT late interaction, and graph context from Graphos, combined into one ranked result.
Mnemo · The memory layer
Permission-aware memory that answers in milliseconds.
Mnemo is Entilla's retrieval and long-term-memory layer. It recalls the right context, for the right person, at the right time, and it never surfaces a document someone isn't allowed to see.
The problem
Enterprise RAG leaks, and it's shallow.
Most retrieval fails on two axes. It ignores permissions, so it can surface restricted documents, and it leans on keyword search, so it misses relationships and time. Employees stop trusting the answers, and security can't sign off on a system that might leak. Both problems kill adoption.
What Mnemo does
Grounded recall, with the ACLs built in.
Mnemo turns a question into cited context by combining GPU-accelerated retrieval over the Graphos graph with a long-term memory store per user and per org. It enforces per-document ACLs at query time, reranks with NeMo Retriever, and returns evidence agents can act on, complete with a confidence score and abstention when it isn't sure.
Core features
What's inside.
Permission-aware
Per-document ACLs enforced at query time, every time.
Episodic memory
Long-term memory per user and per org.
NeMo Retriever
GPU embedding and reranking for high-recall results.
Confidence and abstention
Scores every answer and declines when the evidence is thin.
Temporal recall
Point-in-time answers that respect what was true when.
Why it needs GPUs
Sub-second recall is a GPU workload.
NeMo Retriever
GPU-accelerated embedding and reranking, enterprise-grade and on-prem capable. This is the core capability.
TensorRT
Low-latency embedder and reranker inference, so ColBERT late interaction over large corpora stays sub-second.
Triton and NIM
Multi-model serving at throughput, packaged as NIM for VPC and on-prem deployments.
Per-customer fine-tuned embedders run on A100, H100, and cost-efficient L40S for reranking at scale.
Defensibility
Retrieval quality that compounds.
Per-customer fine-tuned embedders plus accumulated relevance feedback keep improving recall. Switching to a generic vector store means re-tuning everything from scratch and losing that feedback history.
Where it sits
The recall layer of the stack.
Mnemo reads the Graphos graph and feeds grounded context to Nexora for action and Verix for verification. It's pure knowledge-systems retrieval, no new category.