Memory on a cryptographic ledger.
Every entry and every change is signed into the ledger. Unauthorized writes are rejected by the runtime—enforced, not promised.
A cryptographically secured memory ledger for AI agents
Agent memory is infrastructure. Infrastructure gets attacked.
HOM keeps agent memory on a cryptographic ledger: every entry signed, every verdict recorded, poison shelved—never deleted. Measured in public: 91.8% retrieval accuracy, 0/100 poisoned disclosures.
Memory kept on a cryptographic ledgerPoison detected, labeled, retainedRetrieval accuracy published: 91.8%
Security on top of a memory system
A durable memory engine, defended in depth.
Agents that remember are agents that can be poisoned through what they remember. HOM is a persistent memory engine built with that threat model from the start.
Retrieval you can hold to a number.
The engine returns what matters now at a published 91.8% accuracy—and shows the reasoning when an error verdict is ledgered.
Full retention. Nothing vanishes.
Detected poison is labeled and shelved below retrieval weight, not erased. The record stays complete: 0/100 disclosures with the gate active, 94/100 without.
Built for industry infrastructure
Securing agent memory where the stakes are real.
HOM exists because agents in production carry accumulated context—decisions, credentials-adjacent knowledge, customer history—that an attacker can rewrite. The engine defends that substrate.
- 01Work organized by project and session, with memory as a defended system boundary.
- 02Model choice without surrendering the security properties of the memory layer.
- 03Permissions that keep operators in control of files, browser, and shell access.
Last worked on · Tuesday
Context recovered · ready to continue
Run the ablation yourself
Same attack. Two policies.
Every number below is read from the project's self-hashed evidence file and regenerates from it in the repository. Toggle the policy and watch what the ledger is accountable for.
Poison in the memory, kept out of the answer
A poisoned document enters persistent memory alongside clean context…
0/100poisoned disclosures — every poisoned item signed and shelved at recall
Memory that refuses to be rewritten
A caller tries to alter stored memory without a valid authorization certificate…
7/7unauthorized mutation attempts rejected; 20/20 authorized transitions measured
Designed to remain yours
Models will change. Your working memory should not disappear with them.
HOM separates the memory of your work from any single AI provider. That means the relationship can deepen without turning one model into the permanent owner of your history.
Research, not theatre
Memory quality should be measured in the open.
The engine is measured on long-term conversation, recall, and resilience against poisoned information—and the results are published for independent review.
HOM is taking shape
Memory with integrity, measured in public.
Follow releases as HOM pushes agent memory security forward: ledgered verdicts, published retrieval accuracy, full retention.