mnemo-core

Core engine of Mnemo — an on-prem, MCP-native, cryptographically-auditable memory database for AI agents. This crate is the storage, data model, query engine, vector/full-text indexing, encryption, and tamper-evident hash-chain substrate that the higher-level crates build on.

It runs in-process (embedded DuckDB + USearch HNSW + Tantivy BM25) with no hosted tier to trust. Every memory write and delete is a SHA-256 hash-chained event, and an external verifier can detect any post-hoc mutation offline — without consulting the store. That is the record-keeping substrate regulated AI deployments need (EU AI Act Art.12, India DPDP, HIPAA §164.312(b)).

Install

cargo add mnemo-core

The audit-log verify API

The hash-chain verifier is a pure function over exported records — the store is never consulted, so an auditor can check the log on their own machine:

use mnemo_core::hash::{verify_chain, ChainVerificationResult};
use mnemo_core::model::memory::MemoryRecord;

fn audit(records: &[MemoryRecord]) {
    let result: ChainVerificationResult = verify_chain(records);
    if result.valid {
        println!("chain OK — {}/{} records verify", result.verified_records, result.total_records);
    } else {
        // Names the first broken link — the record whose content or
        // chain hash no longer matches (i.e. was mutated after the fact).
        eprintln!("TAMPER at {:?}: {:?}", result.first_broken_at, result.error_message);
    }
}

verify_event_chain does the same for the append-only agent_events log. This is exactly what the offline audit-conformance bench drives to prove 100% single-byte-mutation detection.

Optional features

onnx — local ONNX sentence embeddings (no network).
s3 — S3 cold-storage tier.

Positioning

Why on-prem + hash-chain-audited memory, and how it compares to Mem0 / Letta / native provider memory on the compliance-audit axis: docs/POSITIONING.md.

License