Long-term memory
for your AI.
CogniWeave indexes your team's documents, emails, and files into a persistent vector memory. Every answer is grounded in a specific passage, with a citation that traces back to the source.
- Parse
- Chunk
- Embed
- Store
- AES-256 encryption at rest
- Tenant-isolated per workspace
- SAML 2.0 SSO
- Append-only audit log
- No model training on your data
Every capability in service of one outcome:
an answer you can verify.
Retrieval is only as good as what precedes it. We built the full pipeline — parse, chunk, embed, store, retrieve — because each stage determines the quality of the next.
Every format, one pipeline.
Emails, PDFs, scans, spreadsheets, images — normalised into clean, structured text.
Read the unreadable.
Image-based pages and handwritten scans become searchable text with layout intact.
Right-sized context.
Documents are split into semantically coherent passages so retrieval stays precise.
Meaning as geometry.
Each chunk becomes a dense vector that captures intent, not just keywords.
Grounded answers with sources.
Ask a question. CogniWeave returns the answer with citations linked to the original chunk.
Is the customer entitled to a refund on order #4821?
Based on the refund policy[1], the customer is eligible for a partial refund because the order was delivered after the agreed window[2]. A similar case was resolved by the support team last month[3].
Every answer traceable.
Lineage from query to vector to source — visible in the workspace.
A pipeline, not a black box.
Five stages between raw files and a cited answer.
Ingest
Drop in emails, PDFs, scans, documents, spreadsheets, and images.
Supported: PDF, DOCX, XLSX, PNG, JPG, EML, MBOX, TXT, MD, HTML.
Parse & chunk
Extract text and layout, run OCR on images, split into coherent passages.
Target chunk size: ~512 tokens, 64-token overlap.
Embed
Each chunk is encoded into a vector that places it in semantic space.
Embedding dimension: 1536.
Store
Vectors land in an indexed long-term store — your data, your boundary.
Index: HNSW. Tenant isolation: per-workspace.
Retrieve & cite
Questions trigger nearest-neighbor retrieval; the model answers with citations.
Target retrieval latency: under 200ms.
Not search. Not storage. Memory.
Keyword search returns the documents that contain your words. CogniWeave returns the passages that address your question — regardless of whether they share a single word with your query.
Search finds keywords.
Keyword search is fast and literal. It works well when you know exactly what you are looking for. Vector retrieval works differently: it finds documents that are semantically related to your query, even when they share no words in common. The distinction matters most when you do not yet know what to search for.
Storage holds files.
A file system organises by location. CogniWeave organises by meaning. The same document surfaces wherever it is relevant — not only in the folder you placed it in. Your knowledge becomes navigable by what it contains, not by where it lives.
Chatbots forget.
Language models do not have persistent memory. CogniWeave gives them one. Every answer draws from an index that grows as your team adds knowledge. Every retrieval is logged. Nothing is inferred without a traceable source.
Meaning in geometry.
Every passage we index becomes a point in a high-dimensional space. Passages that address the same idea sit close together, regardless of the words they use.
When you ask a question, retrieval is a traversal of that geometry — finding the nearest neighbours by meaning, not by lexical overlap. The practical result: your question surfaces relevant material even when it was written differently from how you asked.
Clustering by meaning
Semantic neighborhoods surface related work automatically.
Drift detection
Surface stale or contradicting evidence as your corpus evolves.
Citation lineage
Every answer carries the chunk, the document, and the version.
Your data stays yours.
Every boundary is explicit. Every access is logged. We do not train on your data.
Your files, vectors, and queries are yours. We process them to deliver the service. We do not use them to train models, and we do not share them with third parties. You can export or delete your workspace at any time.
Read the security overviewTLS 1.2+ on every request.
AES-256 on stored vectors and source files.
SAML 2.0 single sign-on.
Append-only event log for every retrieval.
Choose where vectors and files live.
In progress — not yet certified.
Dedicated deployment available on request.
Export or delete at any time. No silent training.
Memory built for long horizons.
Most AI memory systems either carry the full history forward and drown in noise, or filter aggressively and pay for it in latency. CogniWeave is built on SimpleMem, an open-source memory architecture from researchers at UNC Chapel Hill and collaborators. It compresses interactions at the point of capture, consolidates related memories over time, and retrieves only what each query needs.
Give your AI a memory.
Vector memory for teams that need answers they can verify and data they can trust.
Connects to your stack
Cogniwave runs on the Model Context Protocol. Any MCP-compatible client, agent, or RPA platform can plug in.
MCP is the open protocol that lets AI clients and tools talk to each other without bespoke adapters.
Cogniwave exposes tools and consumes them. Use it as a host, a client, or both.
Bring your own model, memory, automation, and cloud. Swap any layer without rebuilding the rest.
Need a provider that is not listed. Cogniwave can connect to any MCP-compatible service. Talk to us.