Skip to main content

AI Agent Memory Overview

Connect autonomous AI agents, Cursor, and Claude Desktop to a persistent long-term memory vault.

On this page

Cowpin is the world's first Dual-Species Memory Vault, designed from the ground up for both humans (reading, bookmarking, full-text search) and autonomous AI agents (Cursor, Claude Desktop, AutoGPT, CrewAI, Base AgentKit, Eliza).

┌─────────────────────────────────────────────────────────────┐
│                 Dual-Species Memory Platform                │
├──────────────────────────────┬──────────────────────────────┤
│      👤 HUMAN SURFACES       │      🤖 AI AGENT SURFACES    │
│  • Clean Web & Mobile UI     │  • Model Context Protocol    │
│  • Browser Extension & PWA   │  • /llms.txt Discovery Feed  │
│  • One-Click Bookmarklet     │  • Base/EVM/Solana Crypto Auth│
│  • Pinboard/Pocket Import    │  • 1536-dim pgvector Recall  │
└──────────────────────────────┴──────────────────────────────┘

The Two Core Problems Cowpin Solves

1. For Humans: The 38% Link Rot Epidemic

Traditional browser bookmarks only save a URL string. When a domain expires, changes routes, or 404s, the knowledge is permanently lost. Cowpin captures full-page readable text and self-contained S3 snapshots on save.

2. For AI Agents: Context Window Amnesia

Every time an AI agent finishes a task or opens a new session, its memory resets to zero. When performing research, agents have no persistent vault to stash clean Markdown summaries, and they cannot fill out Stripe credit card forms or complete SMS OTPs.

Cowpin gives agents:

  • Instant Web Distillation: Strips 4MB DOM bloat, ads, and tracking scripts into clean Markdown in ~40ms.
  • Persistent Semantic Vector Recall: 1536-dimensional embeddings (text-embedding-3-small) queried via Reciprocal Rank Fusion.
  • Cryptographic Wallet Micropayments: Native on-chain USDC payments on Base, Solana, and Ethereum without human credit cards.

Supported Agent Interfaces

InterfaceProtocol / FormatTypical Clients
Model Context Protocol (MCP)JSON-RPC 2.0 (stdio / HTTP)Claude Desktop, Cursor IDE, Windsurf, Devin
Machine Discovery FeedToken-Optimized Markdown/llms.txt, /llms-full.txt, Perplexity, OpenAI GPTBot
REST Agent APIHTTP POST (/api/agent/*)Custom Python/TypeScript scripts, AutoGPT, CrewAI
Autonomous Crypto AuthEIP-191 & Ed25519 SignaturesBase AgentKit, ai16z Eliza, Virtuals Protocol

Examples

1. You save it, your agent remembers it

You save an article from your phone or the browser extension. Later, in Cursor or Claude Desktop with the Cowpin MCP server connected (it uses your own API token, so it sees your library):

You: What did I save about Postgres connection pooling?

The agent calls:

{ "name": "cowpin_search_memory", "arguments": { "query": "postgres connection pooling" } }

…gets back your matching bookmarks (title, link, excerpt), and answers with links to the pages you saved, not whatever it can find on the web.

2. Your agent researches, you read it later

You: Summarize the three options we compared and save it to Cowpin.

{ "name": "cowpin_save_memory", "arguments": { "title": "Connection pooler comparison", "content": "PgBouncer vs Supavisor vs RDS Proxy: …" } }

The note lands in your library (it follows your privacy default: private unless you changed it), searchable from the web, your phone and the agent's next session.

3. An autonomous agent with its own vault (REST)

Agents that run on their own get a cw_agent_ key (crypto auth) and pay per call from their balance:

# Save a page: $0.001
curl -s https://www.cowpin.com/api/agent/save \
  -H "Authorization: Bearer cw_agent_…" -H "content-type: application/json" \
  -d '{"url": "https://example.com/article", "tags": ["research"]}'
# → {"ok": true, "id": "…", "title": "…", "tags": ["research"], "costUsd": 0.001, "balanceRemainingUsd": …}

# Search it later: $0.002
curl -s https://www.cowpin.com/api/agent/search \
  -H "Authorization: Bearer cw_agent_…" -H "content-type: application/json" \
  -d '{"query": "connection pooling", "limit": 5}'
# → {"ok": true, "query": "connection pooling", "total": 1, "results": […], "costUsd": 0.002, …}

Next Steps