Verifiable memory for AI agents — 0% hallucination, cited, provably forgettable, deterministic. MCP server. CPU-only, no GPU, zero dependencies.
Server has 13 well-intentioned tools with explicit schemas and descriptions. Naming is verb-forward (learn_, recall, update_, forget, prove_, verify_). Descriptions are detailed and domain-specific (verifiable memory, Merkle proofs, GDPR compliance). However, several parameter descriptions lack specificity (e.g., 'The subject entity' without format/constraint guidance). Output schemas are not formally documented, responses are inferred from code. Some tools mix concerns (recall returns both answer and receipt; prove_fact couples proof generation with fact lookup). Error handling is minimal, no recovery guidance or actionable error messages visible. Parameter validation rules not stated in descriptions. Overall solid for a specialized domain but falls short of production-grade polish.
All exact fact-paths between start and end entities (each path citable).
Audit knowledge for conflicts: for functional relations (one value expected, e.g. 'capital','birthdate') return any (subject,relation) holding >1 live value, showing BOTH sources — call before trusting facts that may have been updated or come from multiple sources.
Permanently and PROVABLY delete a stored fact (GDPR / right-to-be-forgotten). Use when the user asks to forget or remove information — the fact is fully erased and you get a signed proof of deletion. object optional (omit to delete all values).
Full life-line of a fact (all versions live+closed, with sources and valid-time). Audit/compliance.
Merkle root committing the entire current knowledge state (one hash).
Store a fact (subject, relation, object) the agent must recall EXACTLY later, with its source. Call whenever the user states a fact, preference, decision, name, number, or rule worth remembering — it persists across sessions and is never silently distorted. Optional valid_from for valid-time.
Output schemas not formally documented. Responses inferred from code (e.g., recall returns {answer, abstain, cited, receipt}; history returns {history}). LLMs cannot plan downstream tool calls without explicit output schema declarations.
Parameter descriptions lack actionable constraints. 'The subject entity' does not specify format, length, character restrictions, or examples. 'The relation' is similarly vague. LLMs cannot validate inputs without explicit constraints.
Error handling absent or minimal. No recovery guidance visible (e.g., 'Fact not found, try recall() with a broader relation' or 'Invalid timestamp, use ISO 8601 format'). Agents cannot self-correct from failures.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Multi-hop chain: start entity + list of relations, follows subject->object each step (exact only, 0% hallucination).
Merkle inclusion proof that a fact is in the knowledge state (without revealing other facts).
Look up a stored fact and return the answer WITH its cited source — or an honest 'unknown'. ALWAYS call this before answering a factual or memory question instead of guessing: it returns nothing rather than hallucinating, and includes a signed, verifiable receipt. Optional as_of for valid-time.
Counts + current knowledge root.
Update a fact's value without retraining: closes live versions (valid_to=t) and opens a new one. History is preserved.
Verify a Merkle inclusion proof (leaf, proof, root).
Verify a signed receipt returned by recall/forget (tamper-evident).
Tool descriptions for knowledge_root and stats are under 50 characters ('Merkle root committing the entire current knowledge state (one hash).' and 'Counts + current knowledge root.'). Too brief to guide LLM selection or explain when to call them.
Composition concern: recall couples answer retrieval with receipt generation and signature verification. prove_fact couples proof generation with fact lookup. Consider splitting into separate tools (get_fact + get_receipt, prove_fact + verify_proof) for better composability.