convez.ai
MCP ServerSegurançaAvançado
Mars-proj

Verifiable Memory para Agentes IA

Mars-proj/verifiable-memory

Memória com 0% alucinação: cita ou abstém-se, nunca fabrica

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server que fornece memória verificável para agentes IA com zero alucinações. Armazena fatos como dados estruturados, responde apenas com citações ou abstém-se honestamente, permite esquecimento criptograficamente comprovável e auditoria total. Ideal para casos de alto risco (legal, financeiro, compliance).

Instalação

$npx convez-ai Mars-proj/verifiable-memory

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Agentes autônomos em alto-risco (legal, financeiro, healthcare)
  • Sistemas de conformidade e auditoria com rastreabilidade total
  • Esquecimento verificável (GDPR) com prova criptográfica
  • Knowledge bases confiáveis com detecção automática de contradições
  • Workflows que exigem determinismo e respostas reproduzíveis

Veredito

Ferramenta sólida e bem arquitetada para quem precisa de garantias criptográficas de veracidade em agentes IA. Benchmarks convincentes (1M fatos, 137k q/s, CPU-only). Escopo honesto: é camada de confiança, não motor de raciocínio — combine com seu LLM.

README

A tradução para o português deste README ainda está na fila.

verifiable-memory

Memory for AI agents that cannot hallucinate. It answers only from stored facts — with the source cited — or it honestly says "I don't know." Every guarantee below is cryptographic or true by construction, not a prompt trick.

hallucination 0% CPU only deps license protocol pypi

An MCP server + Python SDK. Plug it into any agent (Claude Desktop/Code, LangChain, custom). The LLM phrases; this layer guarantees the facts.


The problem

LLMs store knowledge in weights. So they hallucinate, can't cite, can't be edited, can't forget, can't be audited. That blocks agents from any high-stakes use — legal, finance, healthcare, compliance, autonomous workflows.

What you get (an LLM cannot do these from its weights)

  • 0% hallucination — exact match only; unknown → honest abstention.
  • Citations — every answer carries its source.
  • Provable forgetting (GDPR / right-to-be-forgotten) — the fact is really gone; signed proof; Merkle root reverts.
  • Valid-time — version a fact; ask "as of date T"; full history.
  • Merkle proofs — commit all knowledge to one hash; prove a fact's inclusion without revealing the rest.
  • Contradiction detection — surfaces conflicting values with both sources instead of silently picking one.
  • Signed receipts + determinism — tamper-evident, same query → same answer.

Benchmark (reproducible — python3 benchmark.py)

Stress-tested to 1,000,000 facts on a 7 GB CPU box, no GPU:

Metricverifiable-memory
Hallucination on adversarial traps0.0%
Accuracy when answered / citations100% / 100%
Query latency (p50 / p99)4.4 µs / 14 µs
Throughput137,000 q/s (16 threads)
Memory~1.2 GB for 1M facts (~1 KB/fact)
Provable forget✅ root reverts

vs a naive "always answer" baseline: 0% vs 100% fabrication on the same traps.

Install

pip install verifiable-memory-mcp
verifiable-memory                     # MCP server over stdio
# from source:
git clone https://github.com/Mars-proj/verifiable-memory && cd verifiable-memory
python3 -m vmem.server

Use from Claude Desktop / Code

{
  "mcpServers": {
    "verifiable-memory": {
      "command": "verifiable-memory",
      "args": [],
      "env": { "VMEM_STATE": "~/.verifiable_memory" }
    }
  }
}

Then your agent can learn_fact, recall (cited or abstains), forget (provably), prove_fact, contradictions, multihop, and more — 13 tools.

How it works (1 line)

Facts are stored as data (subject, relation, object + source), indexed for O(1) exact recall; answers are exact-match-or-abstain; the knowledge state commits to a Merkle root. No vectors needed for the verifiable path → 0 fabrication by construction.

Honest scope

This is a memory / trust layer, not a reasoning engine and not a better chatbot. It wins on verifiability (cite-or-abstain, forget, determinism, audit), not on open-ended fluency. Pair it with your LLM: LLM = language, this = ground truth.


🤝 Using this in production?

Need a hosted API, on-prem deployment, or help integrating verifiable memory into your agent (legal / fintech / healthcare / agent platforms)? → Pilot & enterprise: Sergey · svobodg@gmail.com

MIT licensed. PRs welcome.

Ver o original no GitHub

Bom para

Relacionados