What MemVault is

MemVault is a universal, vendor-neutral floating memory layer for your AI agents. Just as SkillVault provides an interoperable home for what your agents know how to do (skills), MemVault provides an interoperable home for what they remember—your architectural decisions, code preferences, past bug investigations, and project constraints.

Crucially, MemVault does not belong to OpenAI, Anthropic, Google, or any single IDE vendor. It is an independent, user-owned cognitive vault reachable from everywhere you interact with AI: in your terminal via Claude Code, in your editor via Cursor or Windsurf, in your browser on ChatGPT.com or Claude.ai, or on your phone talking to Kimi, Gemini, or Claude mobile.

One persistent memory graph. Every platform, every device, every account.

The problem we are solving: The Memory Walled Garden

Agent memory itself has been built inside closed silos. OpenAI has memory inside ChatGPT. Claude Code has memory inside local .claude files. Cursor has its own workspace notes. Each vendor builds a proprietary island.

The real problem is the total lack of interoperability across the tools developers actually use:

Platform Walled Gardens: The memory you build inside ChatGPT on the web is completely trapped inside ChatGPT. It cannot inform Claude Code running in your terminal, and Cursor has no idea it exists.
Cross-Device Amnesia: If you brainstorm an architectural refactor with Kimi or Claude on your phone during your commute, that context is dead when you sit down at your desktop workstation. You are forced to re-explain the entire decision from scratch.
Account & Harness Disconnect: When you switch from Cursor to Windsurf, or from Claude Code to Codex, all historical context resets to zero. Your agent has amnesia simply because you switched windows.
The Human Copy-Paste Tax: Because AI tools don't share state, the human developer becomes an exhausted copy-paste bridge—re-typing guidelines, stack decisions, and bug workarounds across four different chat interfaces.
Proprietary Lock-in: Storing your memory inside a model vendor's proprietary system locks you to their subscription. When a superior model drops tomorrow, switching means abandoning months of accumulated institutional context.

MemVault turns memory into an open, universal utility. It is not tied to any single agent; it is a floating memory layer that travels with you across accounts, platforms, and devices.

How cross-platform memory works: One brain, every surface

Imagine this daily workflow with MemVault:

Morning on Mobile (Kimi / Claude iOS / ChatGPT): On your commute, you discuss a database refactor with Kimi on your phone:
At Your Desk (Cursor / Claude Code / DeepSeek Harness): You sit down at your workstation, launch Cursor, and prompt: "Create the public order route." Your desktop agent performs a pre-flight recall against MemVault, retrieves the ULID decision made on your phone an hour earlier, and implements ULIDs immediately—without you ever typing the requirement twice.
In the Terminal (Git Milestones): Later, you commit code via your terminal. The skillvault git hook ambiently captures the commit summary, diff, and author into MemVault as a durable milestone.
In the Web Browser (ChatGPT.com / Claude.ai): Days later, an engineer asks ChatGPT.com about your API conventions. ChatGPT queries your vault URL and recites the exact ULID convention and the git commit where it landed.

The Four Delivery Surfaces

MemVault makes this possible by delivering your memory across four interoperable surfaces:

1. Authenticated Vault URLs (<code>/v/&lt;user&gt;/t/&lt;token&gt;</code>): Paste or link your vault into any web or mobile AI that can fetch URLs or browse the web—ChatGPT.com, Kimi, Claude.ai, Perplexity, or custom mobile shortcuts. Clean Markdown formatted for LLM attention.
2. Model Context Protocol (MCP): Connects natively into Cursor, Windsurf, Claude Desktop, and ChatGPT Custom Connectors as first-class tools (memvault_recall, memvault_ingest, memvault_session_wrap).
3. Local CLI Tool (<code>skillvault mem</code>): Seamless terminal commands for Claude Code, DeepSeek Harness, Aider, OpenCode, Codex, and local automation scripts.
4. REST JSON API: Standard Bearer-authenticated JSON endpoints (/api/memvault/v1/<user>/recall, /ingest/text, /session-wrap) for custom agent harnesses, CI/CD runners, and internal developer platforms.

Intelligent Cognitive Architecture: Beyond Dumb Text Dumps

Interoperability requires precision. Dumping thousands of words of conversational transcripts into every prompt blows token budgets, drives up API costs, and dilutes model attention.

MemVault is engineered as a high-precision cognitive coprocessor:

### 1. Vector Semantic & Associative Graph Indexing MemVault embeds memories using dense high-dimensional vectors paired with associative knowledge graphs. An agent asking about "ID generation strategy" finds your decision on "ULIDs instead of UUIDv4", even with zero keyword overlap.

### 2. Mathematical Ebbinghaus Decay Dynamics Human memory naturally prioritizes recent, reinforced thoughts over old noise. MemVault applies an Ebbinghaus decay curve: transient debugging scratchpads naturally fade over time unless reinforced through repeated recall (the "savings effect").

### 3. Pinned Golden Rules (Strength 1.0) Critical architectural standards, security rules, and user preferences can be pinned. Pinned memories are mathematically exempt from decay—their strength remains locked at 1.0 forever, ensuring vital guardrails are never forgotten regardless of how much time passes.

[ASKER CONTEXT]: Pinned user preferences and immutable architectural guardrails.
[SITUATIONAL CONTEXT]: Active branch context, recent milestones, and repository trajectory.
[REFERENCE KNOWLEDGE]: Semantically relevant past solutions and bug root causes.

PAE guarantees that floating memory never consumes more than a calibrated fraction of the context window, leaving maximum room for your code and current instructions.

Team & Organization Governance

The memory fragmentation problem is multiplied ten-fold across engineering teams. When one developer solves an obscure Kubernetes DNS race condition or sets up a Stripe webhook pattern, that knowledge remains trapped on their machine.

MemVault bridges the entire team across a Dual-Layered Architecture:

Personal Vault Layer: Individual developers keep private notes, scratchpad priors, and local preferences that never pollute the shared team repository.
Shared Team Vault Layer: Team repositories automatically pool project milestones, architectural decisions, and bugfixes into a centralized organization memory graph.
📌 Pinned Org Standards: Architects and Leads pin golden rules (e.g. "All server routes must authenticate via JWT middleware before touching the database"). These standards are locked at strength 1.0 and injected into the prompt context of every engineer and AI agent across the company.
Contributor Attribution: Every team memory records author attribution (e.g. by @sarah or by @ci-bot), so teams always know who established a pattern and why.
Compliance & Security Controls: Built-in repository blacklists, project-level exclusions, and automated 30, 60, or 90-day retention purge windows ensure strict compliance with SOC2, ISO 27001, and HIPAA policies.

Data Ownership & Vendor Independence

When your agent memory lives inside a closed ecosystem, you are at the mercy of that vendor. If OpenAI changes their terms, Anthropic limits your rate, or a superior open-weight model like DeepSeek or Llama arrives, you cannot take your memories with you.

You Own Your Cognitive Context: Your memory lives in your encrypted vault, not inside a closed model subscription.
Zero Model Training: We never use your memories, commit diffs, or architecture specs to train proprietary or public models.
Switch Models Instantly: Switch from Claude 3.7 Sonnet to GPT-4o to Kimi to an open-source local model running via Ollama without losing a single decision, preference, or architectural milestone.

Pricing

MemVault is built directly into all SkillVault subscription tiers:

Free: 50 persistent memories, 3 skills, 100 MB storage. Perfect for trying universal cross-platform memory on personal side projects.
Pro: €8.95/month. 5,000 persistent memories, 50 skills, 2 GB storage, automatic git commit capture. Ideal for professional developers using AI daily across multiple devices.
Power: €24.95/month. 50,000 persistent memories, unlimited skills, 15 GB storage, neural cross-encoder reranking, and associative graph traversal.
Team & Enterprise: €14.95/seat/month. 100,000 memories per organization, shared team vaults, Pinned Org Standards, RBAC, repository blacklisting, and compliance retention policies.

How to start

Get started in under two minutes:

Create your account at /auth/signup and copy your vault address (https://app.skillvault.dev/v/<username>).
Run skillvault setup --vault-url="https://app.skillvault.dev/v/<username>" on your development machines.
Add your vault URL or MCP endpoint to your mobile AI apps and web assistants. Your agent memory is now truly universal.

Get in touch

Have questions about MemVault, enterprise deployments, or custom compliance workflows? Contact us anytime at support@skillvault.dev.