MemoryLake
MemoryLake is a persistent, multimodal memory layer that plugs into ChatGPT, Claude, Gemini, Qwen and any MCP client, storing conversations, documents, spreadsheets, images and audio so every AI you use recalls your context.
What is MemoryLake?
MemoryLake is a memory layer for artificial intelligence, published by DataCloud Technology Pte. Ltd., a company incorporated in Singapore. Its premise is that the model is interchangeable but what you have accumulated is not, so the memory should live outside the model rather than inside it. The same store is meant to serve ChatGPT, Claude, Gemini, Qwen, Perplexity, OpenClaw, AutoGPT, Manus and anything reachable through an API or the Model Context Protocol.
What it holds is deliberately broad: conversations, plain text, documents, spreadsheets, images, audio and video. These are sorted into six memory types — background (your values and world model, read-only), fact (verifiable, versioned, traceable), event (a chronology), dialogue (compressed, searchable conversations), reflection (patterns the AI has noticed) and skill (methods built once and reused). Two products sit on top: Memory Passport for individuals and Workbrain for organisations, with a component called MemoryLake Router alongside them.
The engineering claims cluster around three ideas. A proprietary engine, MemoryLake-D1, uses visual and logical dual validation to parse complex Excel and PDF layouts. A conflict engine spots contradictions the moment new memories arrive — logical conflicts, implicit contradictions, hallucination conflicts — and resolves them by recency, source priority, confidence weighting, manual review or rules you define, preserving the previous value in version history. And everything is traceable: source attribution, cryptographic integrity verification, exportable audit reports, plus git-like versioning with branches, diffs and rollback.
Open datasets come bundled — PubMed, arXiv and bioRxiv, EDGAR filings, ClinicalTrials.gov, drug databases, USPTO patents, FRED, BLS and World Bank series — alongside connectors for MySQL, PostgreSQL, Delta Lake, Hudi, Iceberg, Office365, Google Workspace, WPS365, Dropbox, Lark and Dingtalk. On security the vendor claims SOC 2 Type II and ISO 27001, AES-256 at rest, TLS 1.3 in transit, optional end-to-end encryption for enterprises, and a triple-party key scheme requiring two shards of three to decrypt. It publishes a self-run LoCoMo benchmark result of 94.03%, with the benchmark code on GitHub. The product is young: the domain dates from September 2025.
What it does
- Store conversations, documents, spreadsheets, images, audio and video in a single memory layer
- Serve that memory to any model or agent through a REST API or a native MCP server
- Detect and resolve contradictions between memories coming from different sources or periods
- Version memory like code, with commits, branches, diffs, rollback and an immutable audit log
- Trace every stored fact back to its source, with timestamps and a full provenance chain
- Query built-in open datasets covering research papers, SEC filings, clinical trials, patents and economic series
- Cut token spend compared with feeding raw files to a model on every call
When to use MemoryLake / When not to
A quick filter to help you decide if MemoryLake is the right fit.
When to use MemoryLake
- Engineering teams whose AI tools forget architectural decisions between sessions — 73 of the 169 published use cases sit here
- Developers juggling several MCP clients such as Claude Code, Cursor, Windsurf, Cline or Zed and wanting one shared memory behind them
- Builders of LLM applications and chatbots that need cross-session context through a documented REST API
- Enterprises trying to keep institutional knowledge alive through staff turnover, which is what the Workbrain product addresses
- Researchers and analysts who can use the built-in open datasets: papers, SEC filings, clinical trials, patents and economic series
When not to use MemoryLake
- Anyone expecting a mobile app — no iOS or Android application is published
- Users looking for a chat assistant: this is a memory layer, not a conversational product
- Teams that require self-hosting or an on-premise deployment, neither of which is offered
- Organisations that need end-to-end encryption or a chosen hosting region on a standard plan, since both are enterprise-only
- Buyers who want a predictable flat fee, because every operation is metered by token
How to use MemoryLake
A typical end-to-end flow, from setup to results.
- Create an account on the MemoryLake console — the free tier needs no credit card
- Decide how you will connect: the native MCP server, the REST API, or a data source connector
- For the MCP route, pick your client among Claude Code, Claude Desktop, Cursor, Windsurf, Cline, Continue.dev and Zed
- Copy the installation command the console gives you and paste the configuration block into that client
- Issue a project-scoped token so access stays limited to one project and can be revoked at any time
- For the API route, work through the reference: authentication, memory operations, workspaces, actors, agents and projects
- Connect the sources you already use — MySQL, PostgreSQL, Delta Lake, Hudi, Iceberg, Office365, Google Workspace, WPS365, Dropbox, Lark or Dingtalk
- Let the platform capture context automatically and load the relevant memory when a session starts
- Review the conflicts the engine flags and choose a resolution strategy, or write your own rules
- Track consumption against the monthly token quota, then add credits or move up a plan when it runs out
Pros & Cons
Pros
- Model-neutral by design: one memory serves ChatGPT, Claude, Gemini, Qwen and any MCP client
- A permanent free plan of 300,000 tokens a month, with no card required
- Per-operation pricing published in full, down to the cost of a single retrieval call
- Portability taken seriously: one-click export, and deletion the vendor describes as irreversible
- Every memory is versioned and traceable to its source, with exportable audit reports
- SOC 2 Type II, ISO 27001, GDPR, CCPA and PDPA claimed, with AES-256 at rest and TLS 1.3 in transit
- Light setup: one configuration block for MCP clients, with no SDK, wrapper or CLI to install
Cons
- No postal address anywhere on the site — the legal identity stops at a company name and a country of incorporation
- The legal and privacy contact sits on a different domain, data.cloud, from the product itself
- Complete silence on whether customer data trains models, and no documented opt-out
- No subprocessor list, and no Article 27 EU representative despite a GDPR claim made from outside the EU
- SOC 2 Type II and ISO 27001 are asserted without a report, certificate number or named auditor
- End-to-end encryption and data residency are reserved for enterprise customers
- Very young: the domain was registered in September 2025 and first archived in February 2026
Pricing & Plans
A permanent free plan is available, offering 300,000 tokens per month with no credit card required. The lowest paid entry point is the Pro plan at USD 19 per month, which includes 6,200,000 tokens; Premium follows at USD 199 per month for 66,000,000 tokens. Annual billing is advertised as saving approximately 17 per cent. Usage beyond a plan quota is drawn from prepaid credits billed from USD 3.125 per million tokens, which do not expire.
- Free — USD 0 per month — 300
- 000 tokens per month — presented as best for trying the product
- Pro — USD 19 per month — 6
- 200
- 000 tokens per month — marked Most Popular
- aimed at regular individual or small team usage
- Premium — USD 199 per month — 66
- 000
- 000 tokens per month — aimed at heavy usage and team-scale workloads
- Credit Balance — prepaid
- from USD 3.125 per million tokens
- no expiry
- auto-recharge available
- consumed only after the plan quota
- Yearly billing on Pro and Premium — advertised at roughly 17 per cent below the monthly rate
Data, GDPR & hosting
A consolidated view of how MemoryLake handles your data.
GDPR overview
Compliance is claimed in explicit terms. The security page states that MemoryLake is fully compliant with the EU General Data Protection Regulation, provides data processing agreements, supports data subject rights, keeps records of processing activities and carries out data protection impact assessments. The privacy policy, last updated in March 2026, lists the full set of rights — access, rectification, erasure, restriction, portability, objection, withdrawal of consent and complaint to a supervisory authority — all exercised through contact@data.cloud. International transfers rely on the European Commission's standard contractual clauses, adequacy decisions or other lawful mechanisms, and customers are notified within 72 hours of a confirmed breach. Two gaps deserve mention: no Article 27 EU representative is named anywhere, although the publisher is established in Singapore, and no data protection officer is identified.
Who owns the data?
The terms are unambiguous: you keep ownership of everything you submit. Section 3 grants DataCloud Technology Pte. Ltd. a worldwide, non-exclusive, royalty-free licence to use, process, store and transmit your content, but only as far as delivering the service requires, and the company explicitly disclaims any ownership claim over user content. In practice the homepage promises a one-click export of everything and deletion described as permanent, with no copies and no backups kept. Access is bounded by a triple-party key scheme — you, the platform and an audited third-party escrow each hold a shard, and two of the three are needed to decrypt — which the vendor presents as making its own reading of your memory architecturally impossible.
Reuse rights
Nothing in the terms limits what you do with your own memories. You keep ownership, you can export the whole store in one click and reuse it wherever you like, which is the stated point of a portable memory passport; the licence runs the other way, from you to the vendor, and is confined to operating the service. On the vendor side, the privacy policy lists service delivery, account management, security, communications, analytics, legal compliance and legitimate interests as its purposes, states plainly that personal data is not sold, and describes sharing only with hosting, analytics, payment and support providers under confidentiality obligations, plus the usual legal and business-transfer cases. Google LLC, through Google Analytics and Google Tag Manager, is the only third party named. One silence is worth flagging: across every page collected, the site never says whether customer content is used to train models, and documents no opt-out.
Data retention & training
Hosting summary
Data residency is offered across nine locations in three regions: Singapore, Tokyo and Sydney in Asia Pacific; Frankfurt, London and Amsterdam in Europe; US East in Virginia, US West in Oregon and Montreal in North America. The choice is made at organisation level and presented as an enterprise capability rather than a default. Once set, memory data, metadata and backups stay within the designated region, and cross-region transfer is disabled unless explicitly configured. Where data does cross borders, the privacy policy relies on the European Commission's standard contractual clauses, adequacy decisions or other lawful mechanisms. The publisher is incorporated in Singapore, so Singapore law and the PDPA apply alongside the GDPR and CCPA commitments made on the security page. Two caveats: the marketing domain itself resolves to an anycast Amazon node in the United States, which says something about the website rather than the platform, and only Singapore is named as a country — the other eight locations are given as cities.
Things to keep in mind
Risks and trade-offs to weigh before adopting MemoryLake.
- Handing one service a complete, cross-tool record of your work and thinking concentrates a great deal in a single place
- A memory that never forgets also keeps your mistakes, abandoned drafts and half-formed opinions, and may resurface them as settled fact
- Leaning on recalled context can quietly erode the habit of restating a problem, which is often how you rethink it
- The vendor says nothing about training models on customer data and documents no opt-out, so sensitive material deserves a decision before upload
- Certifications are claimed but no report, certificate number or auditor is published to back them
- Token metering ties cost to how much you think out loud: an intense month is a more expensive one
- Contractual liability is capped at one hundred Singapore dollars, and no postal address is published to write to
Setup & Integrations
Technical difficulty
Low if you already use an AI client, moderate if you are wiring it into your own application. The MCP route is a single configuration block pasted into Claude Code, Cursor, Claude Desktop, Windsurf, Cline, Continue.dev or Zed; the site claims under two minutes, with no SDK, wrapper or CLI to install. The API route needs a key and a read of the reference: authentication, memory operations, workspaces, agents, projects, error codes and rate limits. A Python SDK covers distributed compute. Nothing runs locally and no infrastructure has to be operated.
Deployment
Integrations
Supported languages
Behind MemoryLake
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement MemoryLake.
Frequently asked questions
Which AI tools does MemoryLake work with?
Is there a free plan?
How does billing work?
What does each operation cost?
Is there an API?
Can the vendor read my memories?
Where is my data hosted?
How long is data kept after I delete my account?
Is customer data used to train models?
Is there a mobile app, and is there an age limit?
Should you pick MemoryLake?
MemoryLake answers a real and widely felt problem: the context you build up with one AI tool evaporates the moment the session ends, and starts from nothing in the next tool. Its response is coherent — keep the memory outside the model, make it multimodal, version it like source code, and let any client read it through an API or a native MCP server. The engineering on show is genuinely dense for a product this young, and the pricing page is unusually candid, publishing the cost of a single retrieval call rather than hiding behind tiers. The permanent free plan makes evaluation cheap.
The reservations are about disclosure rather than capability. The publisher, DataCloud Technology Pte. Ltd., names its country of incorporation but publishes no postal address, and routes legal and privacy contact through a different domain. SOC 2 Type II and ISO 27001 are asserted with no report, certificate number or auditor named. Most notably for a product built to accumulate everything you know, the site never states whether customer content is used to train models, and documents no way to opt out. No subprocessor list is published, and no Article 27 EU representative is designated despite an explicit GDPR claim from a company based outside the EU. The headline performance figures animate on screen and are absent from the underlying page, so they cannot be checked.
Worth trying if you work across several AI tools and the loss of context genuinely costs you time — the free tier asks nothing of you. Worth a procurement conversation, and some direct questions about training and certification, before it becomes the single store of everything your team knows.
- Choosing a selection results in a full page refresh.
- Opens in a new window.