Amnesia by default
Users repeat themselves. Agents re-derive the same conclusion they reached yesterday. Quality resets to zero at the top of every thread.
BetaShared memory across agents is live
Give your AI a memory that outlives the context window
Mnemo is a drop-in memory layer for agents and LLM apps. Store documents, conversations and web pages with a single call, then recall exactly the right context in 41 milliseconds.
- 10,000 memories free
- No credit card
- MCP-native
POST api.mnemo.dev/v1/recall200 OK · 38ms
Powering memory for engineering teams at
Helm SupportVerbalNorthbeam LabsCorticonFieldnotePerch AILoomstackRunway Ops
The problem
Your model meets every user as a stranger
A context window is rented, not owned. The moment the session ends, everything your product learned disappears, and your team rebuilds the same retrieval stack for the third time this year.
Glue nobody owns
Chunkers, embedders, a reranker, a nightly sync job and a vector store, stitched together by scripts that exactly one engineer understands.
Stale facts, stated confidently
Nothing ever expires. Last quarter's pricing and a departed teammate's job title keep surfacing months after they stopped being true.
Without a memory layer
- Context rebuilt from scratch on every request
- Six services to run before retrieval works
- No idea which chunk produced which claim
- Deleting a user means five manual cleanups
With Mnemo
- One scope, one call, memory already warm
- A single endpoint with the pipeline behind it
- Every answer carries its source and timestamp
- One forget call clears the user everywhere
Product
One layer that handles everything between raw data and a good answer
Ingestion, ranking, expiry, provenance and access control are not side quests. Mnemo treats them as the product so your team can stay focused on what it is actually building.
42 parsers built in
Ingest anything
PDFs, Notion pages, Slack threads, transcripts, screenshots, raw text and URLs all go through one endpoint.
Reranked top-k
Recall that ranks
Hybrid vector search, keyword matching and recency weighting run through a cross-encoder so the first result is the one your model needed.
Self-pruning index
Forgetting, on purpose
Deduplication, confidence decay and conflict resolution stop stale facts from quietly poisoning every future prompt.
Scoped namespaces
Shared across agents
One memory pool, scoped per user, per team or per agent. Your research agent inherits what your support agent already learned.
One-line setup
MCP-native
Point any MCP client at your Mnemo server and it gains persistent memory through a single line of config. No adapter to maintain.
No lock-in
Yours to keep
Bring your own vector store, self-host the entire control plane, or export every memory and embedding as JSONL whenever you want.
How it works
Three calls between raw information and a model that knows things
Everything in between is handled for you and stays inspectable, so you can see exactly which memory shaped which answer.
POST /v1/memories4 queued
- notionproduct-spec.md14 chunks
- slack#eng-billing6 chunks
- pdfvendor-agreement.pdfparsing
- urlacme.com/changelogqueued
chunk · embed · extract entities · dedupe
Production ready
Fast enough to sit inside the request, strict enough to sit next to customer data
Memory only helps if it answers before the model needs it and never becomes the reason a compliance review stalls.
Built for the request path
Recall sits inside your model call, not beside it. Warm shards, cached embeddings and a co-located reranker keep p50 retrieval at 41ms from anywhere in the world.
41ms p50 recallEvery answer carries its source
Each memory keeps its origin, author, timestamp and the exact span it came from, so you can render citations and audit precisely why the model said what it said.
Full provenancePrivate by construction
PII is detected and redacted before embedding, memories are encrypted per tenant, and a single delete call removes a user everywhere, including the vector index.
SOC 2 Type II- p50 recall latency
- 41msp50 recall latencyMedian end-to-end retrieval, reranking included.
- memories stored
- 12.4Bmemories storedAcross production workloads on the platform today.
- rolling 90-day uptime
- 99.98%rolling 90-day uptimeMeasured against the public recall endpoint.
- fewer wrong answers
- 38%fewer wrong answersAverage drop in hallucinated facts after switching.
Integrations
It already speaks to the stack you picked
Mnemo is a plain HTTP API with first-party SDKs, framework adapters and a hosted MCP server, so nothing about your architecture has to change to give it memory.
Model
- OpenAI
- Anthropic
- Mistral
Framework
- LangChain
- LlamaIndex
- Vercel AI SDK
Agents
- CrewAI
MCP client
- Cursor
- Claude Desktop
Source
- Notion
- Slack
- Linear
- Google Drive
- Gmail
Store
- Postgres
- Pinecone
- Qdrant
- Supabase
Using something that is not on this list? The REST API is six endpoints and the schema is public, so a new adapter usually takes an afternoon.
Browse the APICustomers
Teams that stopped rebuilding retrieval
What changed for the engineers who moved their memory layer onto Mnemo.
We deleted about nine hundred lines of homemade retrieval glue and replaced it with three Mnemo calls. Our support agent now remembers a customer across email, chat and phone without us stitching anything together.
Retrieval latency was the thing killing our voice agent. Mnemo returns cited context in under fifty milliseconds, so the model never stalls halfway through a sentence.
The forgetting is what sold us. Outdated facts used to leak into answers for weeks. Now a contradiction gets resolved the moment it is written, and we can see exactly which memory won.
Pricing
Priced by recalls, not by seats
Start free with ten thousand memories. Move up only when your traffic does, and never pay separately for ingestion, reranking or provenance lookups.
Hobby
$0forever
Everything you need to ship a side project with real memory.
10,000 memories50,000 recalls per month
- Hybrid vector and keyword recall
- All 42 ingestion parsers
- TypeScript, Python and REST SDKs
- MCP server included
- Community support
Pro
$49per month
For products in production with real users and real traffic.
1 million memories2 million recalls per month
- Everything in Hobby
- Cross-encoder reranking
- Conflict resolution and decay tuning
- Unlimited projects and namespaces
- 99.9% uptime commitment
- Email support within one business day
Scale
$299per month
Dedicated capacity for high-volume agent fleets.
25 million memories20 million recalls per month
- Everything in Pro
- Dedicated index and warm shards
- Bring your own vector store
- SSO and SAML
- Audit log export
- 99.99% uptime commitment
Enterprise
Customannual
Run the whole memory plane inside your own perimeter.
Unlimited memoriesUnlimited recalls
- Everything in Scale
- Self-hosted or VPC deployment
- SOC 2 Type II report and DPA
- Custom data residency
- Named solutions engineer
- Priority incident response
FAQ
Questions engineers ask before they wire it in
A vector database stores embeddings and hands back nearest neighbours. Mnemo handles everything on either side of that: chunking, entity extraction, deduplication, recency weighting, reranking, conflict resolution and provenance. If you already run a vector store you can keep it and point Mnemo at it.
Private beta
Ship an assistant that remembers by Friday
Tell us what you are building and we will send an API key, 10,000 free memories and a quickstart that takes about four minutes.
- · keys issued in batches every weekday
- · no credit card, no sales call
- · export or delete everything whenever you want
