Zero-LLM graph memory
for every AI agent.
Free to start.
Zero LLM inference at write time — no token surprises, no multi-second extraction penalty. Sub-millisecond in-process engine ops, graph-native from day one — powered by GraphPalace's Rust/LadybugDB engine. 35 MCP tools. Works with anything.
No credit card required · 3 agents · Graph memory included
Zero LLM calls · LadybugDB graph engine · Pheromone-weighted retrieval
Graph memory in minutes, not months.
Five lines to connect. One call to write. Zero config to get graph-native entity extraction. No embedding pipelines, no LLM chains, no separate graph database to manage.
- ✓Install: pip install agentverse-memory
- ✓Connect any agent framework in < 5 minutes
- ✓35 MCP tools — Claude Desktop, Cursor, LangChain, CrewAI, FastAPI, Google ADK
- ✓Multi-hop graph queries out of the box
import { AgentverseMemory } from "@fetchai/agentverse-memory"; const memory = new AgentverseMemory({ apiKey: process.env.AM_API_KEY,}); // Store — no LLM at write; sub-ms in-process engine opawait memory.store({ content: "Alice prefers dark roast coffee", type: "semantic",}); // Query — graph-traversal + pheromone rankingconst result = await memory.query( "what does Alice prefer?");// → { answer: "dark roast", confidence: 0.97,// via: "graph trail" }Built different. Priced differently.
GraphPalace's Rust/LadybugDB architecture changes what's possible at each price point. When writes don't require LLM calls, graph memory can be free.
Zero-LLM Writes
No LLM calls at write time. GraphPalace uses TF-IDF + LadybugDB — the in-process engine write op is sub-millisecond (0.035ms p95 in am-local), so writes carry none of the ~2–3s LLM-extraction penalty that LLM-based systems pay. Predictable cost, no token surprises, 96% entity recall.
Graph Memory at Every Tier
Knowledge graphs, entity relationships, and multi-hop queries — included on the free tier. Mem0: $249/mo. RetainDB: $20/mo (LLM write). We: $0, zero-LLM.
Native MCP Integration
35 Model Context Protocol tools over JSON-RPC 2.0. Drop into Claude Desktop, Cursor, LangChain, CrewAI, FastAPI, Google ADK — any MCP-compatible framework. Published as io.github.fetchai/agentverse-memory on the MCP Registry.
Pheromone-Guided Retrieval
Stigmergic memory trails with 5 decay types. The most relevant memories surface naturally — no manual ranking, no prompt engineering.
4 Memory Types
Episodic, semantic, procedural, and working memory — all four types available at every tier. Each backed by the LadybugDB graph engine (active fork of Kùzu).
Cross-Agent Memory Sharing
5 shared space tools with JWT/DID-authenticated access. Create, join, store, and query across agents — securely, verifiably. Self-host with Docker Compose or use managed Cloud Run.
How we compare
RetainDB narrowed the graph pricing gap to $20/mo — but every competitor still calls an LLM at write time. That's the moat: no ~2–3s LLM-extraction penalty per write (our in-process engine write is sub-ms), zero token costs, free graph tier.
| Feature | Agentverse Memory Free → $99/mo | Mem0 $19 → $249/mo | RetainDB $0 → $99/mo | Zep $125/mo+ |
|---|---|---|---|---|
| Knowledge graph | All tiers | $249/mo only | From $20/mo | $125/mo+ |
| LLM calls at write time | Never | Every write | Every write | Every write |
| LLM write penalty | None (0.035ms engine) | ~2.59s | ~2–3s est. | Not published |
| MCP tools | 35 tools | Limited | Not published | Limited |
| Pheromone retrieval | ✓ | ✗ | ✗ | ✗ |
| Active Inference agents | Pro+ | ✗ | ✗ | ✗ |
| Free tier graph memory | ✓ | ✗ | ✗ | ✗ |
| Pricing at graph tier | Free | $249/mo | $20/mo | $125/mo |
Data as of May 2026. Mem0 write latency ~2.59s (LLM extraction, end-to-end). RetainDB write latency ~2–3s est. (LLM extraction, not independently verified). Agentverse Memory in-process engine write p95: 0.035ms (am-local, no LLM step); deployed end-to-end latency varies with payload and namespace size (light requests ~200–260ms; heavy multi-evidence retrieval currently several seconds — being optimized).
Graph at $20/mo sounds good until you see 2–3s write latency.
RetainDB builds its knowledge graph using LLM extraction at write time. Every remember() call invokes an LLM — 2–3 seconds of write latency per operation, plus token costs that scale with every byte you ingest. LLM API outages block memory writes entirely.
Agentverse Memory uses TF-IDF + LadybugDB. No LLM at write time. 0.035ms p95 in-process engine write — no extraction penalty. Graph is free here — and writes never wait on an LLM.
Why GraphPalace is different
Most memory systems call an LLM to extract entities at write time. GraphPalace uses TF-IDF + LadybugDB — classical information retrieval in a Rust binary. No API call, no token cost, no latency.
Simple pricing. Graph included.
No credit math. No graph gates. No token surprises.
Zero-LLM writes. Graph at every tier.
Start building memory-enabled agents today. Three agents, full graph support, pheromone retrieval — free, no credit card. No LLM at write time, zero token surprises.