AI Agent Memory API — Persistent State, Free

Every AI agent session starts with a blank slate. Your agent learns something, computes something, decides something — and then the session ends and it's all gone. The next run, it starts over. Here's how to fix that without provisioning a database, configuring Redis, or writing a single line of infrastructure code.

Why agents forget

Context windows are not memory. They're working memory for one session. When the session ends, everything in the context window evaporates. This causes problems that compound over time:

• Agents re-fetch data they've already retrieved
• Personalization resets on every run — no learning curve
• Multi-step workflows can't pass state between sessions
• Multi-agent systems can't share context without a shared database
• Debugging is hard because there's no history of what the agent actually did

The obvious fix is Redis or DynamoDB. But you have to provision it, configure access, write read/write code, handle TTL, and maintain it. For an agent prototype, that's a full infrastructure decision before you've validated anything.

Three memory layers

Slopshop provides three distinct memory primitives, each suited to different patterns:

Key-value: store and retrieve state

The simplest case. Store anything as JSON under a key, retrieve it in any future session, on any machine, with any LLM, as long as you use the same API key.

curl -X POST https://slopshop.gg/v1/memory-set \
  -H "Authorization: Bearer demo_key_slopshop" \
  -H "Content-Type: application/json" \
  -d '{
    "key": "agent:research:transformer-papers",
    "value": {
      "papers": ["arxiv:2401.00001", "arxiv:2401.00042"],
      "summary": "Key insight: sparse attention reduces quadratic complexity",
      "fetched_at": "2026-03-26T10:00:00Z",
      "confidence": 0.88
    },
    "ttl": 604800
  }'

# Response: { "ok": true, "stored": true, "credits_used": 0 }

curl -X POST https://slopshop.gg/v1/memory-get \
  -H "Authorization: Bearer demo_key_slopshop" \
  -d '{"key": "agent:research:transformer-papers"}'

# Returns the full value you stored, including metadata.
# The agent using GPT-4o today gets what the Claude agent
# stored last week. Cross-LLM memory, zero configuration.

Knowledge graph: store what your agent learns

When your agent is researching, it's not just accumulating facts — it's building a web of entities and relationships. A knowledge graph is the right structure for that. Store Person, Company, Concept, Paper nodes and the typed edges between them.

curl -X POST https://slopshop.gg/v1/graph-add \
  -H "Authorization: Bearer demo_key_slopshop" \
  -d '{
    "nodes": [
      {"id": "vaswani_2017", "type": "Paper", "props": {"title": "Attention Is All You Need", "year": 2017}},
      {"id": "concept_attention", "type": "Concept", "props": {"name": "Self-Attention Mechanism"}}
    ],
    "edges": [
      {"from": "vaswani_2017", "to": "concept_attention", "type": "INTRODUCES", "props": {"significance": "foundational"}}
    ]
  }'

curl -X POST https://slopshop.gg/v1/graph-query \
  -H "Authorization: Bearer demo_key_slopshop" \
  -d '{
    "start": "concept_attention",
    "direction": "incoming",
    "edge_type": "INTRODUCES",
    "depth": 2
  }'

# Returns: all papers that introduced or built on self-attention,
# up to 2 hops from the concept node. The graph grows each
# session as the agent discovers new relationships.

Cross-LLM, cross-device, cross-session

Memory in Slopshop is keyed to your API key, not to a model or a session. That means:

• A Claude agent stores a result on Tuesday. A GPT-4o agent retrieves it on Friday. Same data, seamlessly.
• An agent running on your laptop stores user preferences. The same agent on a production server reads them. No sync required.
• You swap models — from Claude 3 to Claude 4, from GPT-4o to Gemini Ultra. The memory doesn't care. It's just key-value data accessible via HTTP.
• Multi-agent pipelines share state without any inter-agent communication infrastructure. Agent A writes, Agent B reads.

Compared to Redis and DynamoDB

Full list of free memory operations

Using memory from Claude Code (MCP)

Add Slopshop to Claude Code's MCP config and all memory tools become native Claude tools. No curl, no HTTP — Claude calls memory_set, memory_get, graph_query the same way it uses any other tool.

{
  "mcpServers": {
    "slopshop": {
      "command": "npx",
      "args": ["-y", "slopshop-mcp"],
      "env": {
        "SLOPSHOP_API_KEY": "your_key_here"
      }
    }
  }
}

// Claude now has memory_set, memory_get, graph_add,
// graph_query, queue_push, queue_pop as native tools.
// 0 credits. No database to maintain.