The Memory That Remembers,
Enforces Your Rules, and Reviews the Work.
Your Ai forgets everything between sessions and repeats the mistake you corrected yesterday. CogmemAi is one memory for every Ai tool you use: it remembers your decisions across sessions, tools and models, saves what matters on its own, stops any action that breaks your rules before it runs, and reviews the finished work against the project's intent.
npm install -g cogmemai-mcp && npx cogmemai-mcp setup
One Memory, Six Jobs
Every new session starts from zero, and built-in memory is a flat file with no search and no judgment. CogmemAi remembers, saves on its own, enforces your rules and reviews the work. Each card opens the full story.
Remembers what matters
Decisions, corrections and context, recalled by meaning across sessions, tools and models. 95.10% on LongMemEval, the top published score.
How the memory engine works →Saves on its own
Decisions, fixes and deployments land in memory without your Ai choosing to save. When Claude Code compacts its context, CogmemAi reloads what was lost.
See it in Claude Code →Stops what breaks your rules
Save a rule in plain words and the next session denies the command before it runs, over ssh too, with your words as the reason.
How Guard works →Rule packs, one command
40 DevSecOps rules, plus AWS, Azure and Google Cloud packs. Shell commands are stopped before they run; plain-English actions are judged.
Read the DevSecOps rules →Reviews the work
Finished work is scored against the project's intent: what was covered, what was missed, any rule broken. The migration playbook gates every phase this way.
The cloud migration playbook →Gives any Ai an identity
Name, voice, values and boundaries that persist across every tool and model, written once and enforced everywhere.
CogmemAi Identity →Inside the Memory Engine
It gets better the more you use it: it learns what helps, connects related knowledge and teaches your Ai how you work.
- Think Before You Speak Your Ai checks memory before suggesting anything already tried or rejected.
- Semantic search + reranking Vector and keyword retrieval, fused, then precision-scored by a neural reranker.
- Self-improving recall Memories that help rise; ones you never use fade and archive on their own.
- Knowledge graph Related memories link automatically when you save them.
- Contradiction detection Conflicting memories are flagged before stale facts reach your code.
- Query synthesis One coherent answer drawn from every relevant memory, not a list of matches.
- Context-aware ranking Debugging, planning or reviewing: the right kind of memory surfaces first.
- Cross-project intelligence Patterns seen across projects are promoted to global scope.
- Auto-skills Corrections become behavioral rules whose confidence rises or falls with results.
- Mandatory rules NEVER and ALWAYS rules load in every session and bypass all scoring and decay.
- Health score A 0 to 100 score with the factors that move it.
- Session replay Every session opens with a summary of the last one.
The full technical breakdown · How we scored 95.10% on LongMemEval · The built-in Claude Skill
Cloud, Local or Hybrid
The same npx cogmemai-mcp setup wizard for all three. Every mode needs a free API key and is encrypted.
Cloud Recommended
The full engine: semantic search, everything above, cross-device sync and team sharing. Nothing runs on your machine but a thin MCP client.
Local
Data never leaves your device. Full-text search, works offline. Upgrade to cloud with one command.
Hybrid
Saves to both, reads from cloud, falls back to local offline and syncs when you reconnect.
Every Tool, Every Model, Every Teammate
A memory saved in Claude Code is there in Cursor a second later, on any model, on a new laptop, and for every teammate's Ai. Sensitive memories stay personal.
CogmemAi for Teams · Agent frameworks · Robots and drones · On-premises (Edge)
Why Cloud Memory Wins
| CogmemAi Cloud | Any Local Memory | Built-in CLAUDE.md | |
|---|---|---|---|
| Benchmark accuracy | 95.10% LongMemEval, 91% LoCoMo | 20-40% typical | N/A |
| Saves without the Ai choosing to | Yes | No | Manual only |
| Compaction recovery | Automatic | None | None |
| Rules enforced before a command runs | Yes, plus rule packs | No | No |
| Search | Vector + keyword + reranking | Keyword or basic vectors | None |
| Other machines, editors and teammates | Yes | One machine | Claude only |
| Quantum-safe encryption | Yes | Varies | Plaintext |
| Price | Free, Pro $9.99/mo | Free (until it breaks) | Free (200 lines) |
CogmemAi vs Mem0 vs Zep · CogmemAi vs MEMORY.md · Ai memory glossary
Setup in 60 Seconds
Zero Install (Remote)
Connect directly, no npm, no Node.js, no config files. Just add the remote endpoint to your MCP client:
Endpoint: https://hifriendbot.com/mcp/
Auth: Bearer cm_YOUR_API_KEYWorks with any MCP client that supports Streamable HTTP transport, Claude Desktop, Cursor, and more. Get your free API key to connect.
Not using MCP? The same engine is available over plain HTTPS from any language: REST API reference. Or install locally for compaction recovery hooks and offline support:
Run the setup wizard
npx cogmemai-mcp setupThe wizard asks you to choose a mode (Cloud, Local, or Hybrid), then walks you through the rest, API key verification, Claude Code configuration, and compaction recovery hooks.
Get your API key (Cloud & Hybrid only)
Sign up free and generate a key on this page. All modes require a free API key, for local mode it works like a software license key, and your data stays on your machine.
Done
Restart Claude Code. It now has persistent memory with automatic compaction recovery, it remembers your architecture, preferences, and decisions across every session. No prompting needed.
Manual setup (if you prefer not to use the wizard)
Per project, create .mcp.json in your project root:
{
"mcpServers": {
"cogmemai": {
"command": "cogmemai-mcp",
"env": {
"COGMEMAI_API_KEY": "cm_your_key_here"
}
}
}
}Or global, works in every project:
claude mcp add cogmemai cogmemai-mcp -e COGMEMAI_API_KEY=cm_your_key_here --scope userLocal mode (free API key required, data stays local):
claude mcp add cogmemai cogmemai-mcp -e COGMEMAI_MODE=local --scope userManual setup does not install compaction recovery hooks. Run npx cogmemai-mcp setup afterward to add them.
Pricing
Start free. Upgrade when you need more.
Personal
- 1,000 memories
- 500 saves/mo
- 10 projects
- Semantic search
- Priority support
Team
- 10,000 memories
- 5,000 saves/mo
- 50 projects
- Shared team memory
- Priority support
Enterprise
- 50,000 memories
- 20,000 saves/mo
- 200 projects
- Shared team memory
- Priority support
Edge (On-Prem)
- Full engine on your hardware
- Air-gap and sovereign deployment
- Your keys, quantum-safe encryption
- Dedicated engineering support
- Annual license
Or pay per operation with USDC on-chain via x402, no credit card or subscription required.
Frequently Asked Questions
Does CogmemAi see my source code?
No. CogmemAi stores extracted facts, short sentences like "This project uses React with TypeScript" or "Auth uses JWT tokens." Your actual source code never leaves your machine.
What happens when Claude Code compacts my context?
CogmemAi detects compaction automatically and saves a summary of your session before it happens. After compaction, your memories are reloaded so you can keep working without re-explaining anything.
Does it work with Cursor, Windsurf, and other editors?
Yes. CogmemAi works with any MCP-compatible Ai system: Claude Code, Cursor, Windsurf, Cline, Continue, autonomous systems, and more. Compaction recovery hooks are specific to Claude Code, but memory storage and retrieval work everywhere.
If I switch editors, do my memories follow?
Yes. Memories created in Claude Code are instantly available in Cursor, Windsurf, Cline, or any other MCP client. Your knowledge lives in the cloud, not in any specific tool.
Does it work across different Ai models?
Yes. CogmemAi is model-agnostic. Switch between Opus, Sonnet, Haiku, or any model your editor supports. Your memories persist regardless of which model is running.
Why not just use a local SQLite file?
CogmemAi offers local mode with full-text search (FTS5) and your data stays on your machine. A free API key is required for registration (like a software license key) but nothing leaves your device. However, local mode doesn’t include the full Intelligence Engine. Cloud mode (which is quantum-safe and superior in every way) adds semantic search (find memories by meaning), auto-linking knowledge graph, contradiction detection, self-improving recall, auto-skills, query synthesis, cross-device sync, and team collaboration. A local-only file also doesn’t follow you to a new machine, a new editor, or a teammate. Upgrading from local to cloud takes one command.
Can I export my memories?
Yes. You can export all your memories as JSON from the dashboard or via the MCP export_memories tool. You can also import memories from a JSON file. Your data is always yours.
Is there a free tier?
Yes. The free tier includes 500 memories, 500 extractions per month, and 5 projects. No credit card required. Upgrade when you need more.
How is my data secured?
All memories are protected with quantum-safe encryption, built to withstand both today's threats and tomorrow's quantum computers. Your data is encrypted before it’s written and decrypted only when recalled. API keys are cryptographically hashed (irreversible). All traffic is encrypted over HTTPS. Your data is never used for model training and is never shared between users. You can delete everything instantly. We store only extracted facts (short sentences), never raw source code.
Can I pay with crypto instead of a credit card?
Yes. Operations beyond the free tier can be paid per-use with USDC on-chain via the x402 protocol, no credit card or subscription required. When your free limits run out, the API returns a 402 response with USDC payment instructions. Pay on-chain, retry your request, and the operation completes. You can also subscribe via Stripe if you prefer automatic billing.
Quantum-Safe, Enterprise-Grade Security
Quantum-safe encryption
Every memory, cloud and local, encrypted to resist classical and quantum attacks.
No source code stored
Only extracted facts, short sentences. Keys are hashed; all traffic is HTTPS.
Delete anytime
Wipe everything from the dashboard or one tool call. Never used for training.
Never shared
No cross-user sharing. Sensitive data is detected and excluded automatically.
Stop Re-Explaining Your Codebase
Join developers who never lose context again.
Also from CogmemAi: CogmemUI, 14 Ai models in one chat with memory · CAT, memory in Gmail, Slack and any site · AiTotalRecall