Memory infrastructurefor AI.
Geniffy is the memory infrastructure for AI applications and agents. It turns every conversation, document and decision into lasting, verifiable knowledge, served to any model in milliseconds.
14 days of Pro free · No card · Export or erase any time
Where each project stands, what was promised and by when, and what happened, on one page.
Works with the models, frameworks and tools you already use
- Claude
- ChatGPT
- Gemini
- Cursor
- VS Code
- LangChain
- CrewAI
- Vercel AI SDK
- n8n
- Zapier
- Make
- Pipecat
- Pydantic AI
- LiveKit
- Haystack
- Perplexity
- Mistral
01The memory layer
Beyond stored facts. The full picture.
Most memory keeps isolated facts and returns the nearest matches. Geniffy reads whole conversations and agent sessions, and keeps what an expert would: where things stand, what was decided, what was promised and what was learned.
AskedWhere does this refund stand?
- Customer since 2024.
- Asked about a refund.
- Prefers email.
Geniffy- NowRefund approved on 14 Oct, waiting on the bank.
- DecidedA replacement instead of store credit, as the customer asked.
- Due FridayThe callback the agent promised.
- ChangedDeliveries now go to the new office address.
Current state. Where each user, account or project stands: what is open, what comes next, and what was decided and why.
Commitments. What was promised, and by when. Marked done when kept, and surfaced before it slips.
Episodes. Whole conversations and agent sessions, tool calls included, kept as one record of what was tried, what failed and what worked.
Lessons. What worked and what did not, kept so the next session does not repeat a mistake.
Consolidation. Every night, conflicting information is reconciled, each change is recorded with its reason, and the memory tests itself.
The briefing. All of it in one call before any model starts. Read, not generated: returned in milliseconds, and free on every plan.
- Your application, through the API and SDKs
- Calls, chats, emails and agent sessions
- Documents, files and web pages
- Claude, ChatGPT, Cursor and coding agents, through MCP
- Current state, commitments, episodes, lessons and facts
- Consolidated every night, with every change recorded
- A briefing before every session
- Answers with their sources, or no answer
- Search, export and erasure
- Served to any model and any framework
Every source, one memory.
Conversations, agent sessions, transcripts, documents and pages.
Start buildingEvery session opens with a briefing: where things stand, what was promised and by when, and what happened, drawn from the user's own conversations.
02Verifiable
Answers with sources. Never guesses.
Each memory keeps the exact words it came from, who said them and when. When nothing supports an answer, Geniffy says so instead of guessing.
Ask the memoryChronology. It knows what was true, and when. A change of plan replaces the earlier fact, which stays on record.
Restraint. No source, no answer. Ask about something never said, and Geniffy says so instead of inventing a reply.
Met the Lumen team today. Priya runs product there, and she wants the pilot criteria before the demo.
Lumen Labs demo prep · 3 OctProvenance. Any memory opens to the exact words it came from, and can be corrected or forgotten.
Benchmarks
Measured on public benchmarks.
Answer accuracy, time-aware questions, false premises and source fidelity, judged with the LongMemEval method.
03For developers
Built for developers.
The same memory, as an API: save what each user says, and serve what matters before every reply, with the source of every line.
1from geniffy import Geniffy
2
3client = Geniffy() # reads GENIFFY_API_KEY
4mem = client.space("customer_1042") # one memory per user
5
6source = mem.memories.add(
7 "Priya Nair signs the Lumen renewal. It moved from March to June."
8)
9mem.sources.wait(source.id) # learning takes a few seconds
10
11print(mem.context("Who signs the renewal, and when?"))
Isolated per user. Every call stays inside the user's own space, and a key can be limited to that user. Nothing else can read it.
Four ways to recall. briefing() to open a session where it left off, context() for your prompt, ask() for an answer in words, search() for the memories.
Export and erasure. Export everything held for a user, or erase it, in one call.
Correct and forgetWorks with your stack.
41 integrations, each tested and documented.
- AnthropicModel
- OpenAIModel
- GeminiModel
- LangChainFramework
- LangGraphFramework
- LLlamaIndexFramework
- OpenAI Agents SDKAgents
- Vercel AI SDKFramework
- CrewAIAgents
- Pydantic AIAgents
- Google ADKAgents
- MMastraAgents
- PipecatVoice
- LiveKitVoice
- HaystackFramework
- n8nAutomation
- ZapierAutomation
- MakeAutomation
- GmailData source
- Google DriveData source
- NotionData source
- SlackChannel
- WhatsAppChannel
- +18More integrationsTwilio, Zendesk, Linear and more
What teams build with Geniffy.
Support that knows the customer.
- An agent that knows the customer.
Their plan, their recent tickets and what they were promised, before the first reply.
- Context across every channel.
What a customer said on chat, email or WhatsApp is known on the next call.
- Handoffs with full history.
A human agent takes over with every fact and the line it came from.
Dana Whitfield · Chat · 09:41
The export is failing again.
mem.briefing(cue="The export is failing again.")4 sources
- NowOpen since 9 Oct: CSV export times out on large workspaces.Ticket · 9 Oct
- DueAn update, promised for 21 Oct, when the fix ships.Email · 12 Oct
- LessonExporting one month at a time works until then.Chat · 10 Oct
- KnownBusiness plan, 40 seats. Prefers email to calls.Call · 2 Sep
Reply
Sorry, Dana. This is the timeout we found on 9 Oct, and the fix ships on 21 Oct. Until then, exporting one month at a time works, as it did last week. I will email you when it is live.
04Security and privacy
Private and secure by default.
Memory is never sold, never shown to advertisers and never used to train a model. Every user's memory is isolated, and can be exported or erased at any time.
Isolated per user. Each user's memory lives in its own space. A key can be limited to one user, and nothing else can read it.
Encrypted throughout. In transit and at rest, with daily backups kept for 35 days in the same region.
Export and erasure. Export everything held for a user, or erase it: removed at once, and from backups within 35 days.
Scoped access. Every AI app connects through OAuth with its own permission, read-only if required, and can be disconnected at any time.
Commitments
What is in place today, and what is on its way. A status changes only once it is true of Geniffy.
Enterprise
Your own contract and a security review.
Built on proven infrastructure.
- Anthropic
- OpenAI
- Google Gemini
- Amazon Web Services
- Google Cloud
What's new.
Every update- When something ends, memory stops serving itAccess removed, a plan cancelled, a rule dropped: once a later word says so, the old memory stops being served and stays as history.
- A saved text comes back word for wordAsk for a saved prompt, rule or document and you get the text itself, page by page, not just the facts learned from it.
- Every save counts what it holdsA save counts the tokens it holds, so a ten-word note is about 15. The minimum of 100 tokens a save is gone.
Start free. Retrieval is never metered.
Compare every plan- Free$0foreverFor individuals trying memory in their own agent or editor
- Unlimited retrieval
- 500K tokens learned a month
- 100 written answers a month
- ProRecommended$20a monthFor developers shipping an agent to real users
- Unlimited retrieval
- 5M tokens learned a month
- 2,500 written answers a month
- Max$100a monthFor products serving thousands of users
- Unlimited retrieval
- 25M tokens learned a month
- 12,500 written answers a month
Build with Geniffy.
Start with 14 days of Pro, free.
- Measured on Geniffy's memory engine test harness in September 2026: LoCoMo, MSC and ABCD conversations and synthetic cases, 190 to 1,373 questions per figure, judged with the LongMemEval judge prompt.