Verifiable Memory

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.

> 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:

| Metric | verifiable-memory | |---|---| | Hallucination on adversarial traps | 0.0% | | Accuracy when answered / citations | 100% / 100% | | Query latency (p50 / p99) | 4.4 µs / 14 µs | | Throughput | 137,000 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.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Mars-proj
• Source: Mars-proj/verifiable-memory
• License: MIT

Install and usage instructions live in the source repository linked above.