Nerve Memory
Memory that knows where to look
Nerve Memory gives Agents and Command the right context, from your documents, from what your team has decided, and from what you chose to remember, and shows where every answer came from.
Nerve Memory · recall
Knowledge
Your documents and policies
Project memory
Decisions for this project
Personal memory
Only you can see it
Live state
Read fresh, never remembered
Demonstration with sample data
Nerve Command
- 1You say it
- 2Command proposes
- 3You decide
Demonstration with sample data
Three tiers of embedding
- Tier 1In useYour provider's embedding model
Connect OpenAI or OpenRouter and Nerve uses their embedding models by default, with similarity thresholds measured for each model.
Personal-sensitivity memories are never sent to a hosted model.
- Tier 2Your own endpoint
Point Nerve at a self-hosted, OpenAI-compatible embeddings endpoint. Text stays on infrastructure you run.
Narrower vectors are padded without changing similarity.
- Tier 3Nerve's own engines
Always there, nothing to configure. Two engines built by Nerve, running inside Nerve, with no key needed and no data leaving your deployment.
The floor under every workspace: Memory and Knowledge work from the first minute.
NerveMemory Engine V3for Memory
nerve-memory-local-v3
The engine behind Nerve Memory. Free, private and in-process, so what your team asks Nerve to remember is recalled without leaving your deployment.
NerveHashing Engine V2for Knowledge
nerve-local-hash-v2
The engine behind Nerve Knowledge. Documents and web pages are searchable the moment they are added, with no key and nothing to set up.
Every stored vector records the model that made it, and Nerve only ever compares vectors from the same model. Change tiers whenever you like: Nerve re-embeds and never mixes models.
- 01Knowledge from your sources
- PDF, Word, text and Markdown, and your own web pages, become passages Agents can cite. Uploads are checked by content, not file name.
- 02Memory you control
- Personal memory is visible only to you. Project memory stays inside that project. Workspace memory is shared deliberately by a manager.
- 03The model proposes, you save
- Memory is created from what you type or a card you press. The model has no way to write it on its own.
- 04Sensitive things refused
- Credentials, key-like tokens, card numbers and similar identifiers are refused by code, not by a polite request to the model.
- 05Forget on request
- Remove any item, or tell Command not to remember a conversation. Forgetting reaches recall as well.
- 06Never authority
- Live state such as connections, runs and approvals is always read fresh. A remembered action is not evidence that it succeeded.
How it is different
Remembering is easy. Remembering safely is the work.
A pile of embeddings can find text. Nerve Memory decides what may be remembered, who may see it and when it must not be trusted.
Who writes memory
A typical memory layer:The model saves whatever it thinks is useful, including things it read on a web page.
Nerve:The model can only propose. A proposal made after reading outside content says so on its card, and you press Save.
Who can see it
A typical memory layer:One shared store, filtered in the app.
Nerve:Scoped by person, project and workspace, and enforced in the database under the caller's authorization.
Finding it
A typical memory layer:Vector search alone, with one global threshold.
Nerve:Keyword and vector recall fused in one query, with similarity floors measured per embedding model.
Stale facts
A typical memory layer:A remembered state is treated as current.
Nerve:Connections, runs and approvals are always read live. A memory of an action is never proof it worked.
Sensitive text
A typical memory layer:Stored, and sent wherever the embedder lives.
Nerve:Secrets and identifiers are refused. Personal-sensitivity text is never shared with the workspace and never sent to a hosted embedder.
Proof
A typical memory layer:An answer, with no way to tell where it came from.
Nerve:Answers cite the exact passage, and Command shows which memory shaped a reply.
A comparison of design approaches, not a benchmark.
Powered by Nerve Cortex. Every Nerve Agent, every Command turn and every workflow step runs on one durable runtime, built by Nerve, that keeps its place, asks before it acts and never repeats an effect.
Meet Nerve CortexDeploy your first Nerve Agent
Start with one conversation your team handles every day. Give the Agent what it needs to know, connect the tools it should use, and decide where a person steps in.