Every memory on the record.
Entries are signed in through a cryptographic gate. Unauthorized writes are rejected by the runtime, not by policy.
The product
Agent memory, defended like infrastructure.
What your agents remember is signed in, judged on the record, and kept in full—accurate when trusted, shelved when poisoned.
A different starting point
Built for agents that operate where trust matters.
HOM treats what an agent remembers as critical infrastructure—signed, judged, and retained in full.
Retrieval held to a published number.
91.8% accuracy on long-horizon recall—measured, published, and reproducible, with error verdicts ledgered alongside their reasoning.
Poison shelved, never erased.
Detected poison is labeled and driven below retrieval weight—out of the agent's context, still on disk for audit and appeal.
Operators decide what agents can touch.
Files, browser, shell, and connected tools are granted deliberately, with boundaries operators can verify.
The standard
Accurate retrieval, on the record.
When an agent answers from memory, the ground truth is the record: what was recalled, what it was labeled, and why. HOM holds that entire chain to a published number.
- 01Retrieval accuracy published and reproducible: 91.8% on long-horizon recall.
- 02Error verdicts ledgered with their reasoning—no silent degradation of what the agent knows.
- 03Full retention: superseded and poisoned entries stay on disk, below retrieval weight.
Signed inEntry joins the ledger
JudgedVerdict recorded with reasoning
RetainedTrusted or shelved—never erased
Provider-agnostic by design
Any intelligence. The same guarantees.
Swap providers or run local models—the ledger, the verdicts, and the retention guarantees do not depend on which model reads the memory.
OpenAIAnthropicOpenRouterOllamaLocal modelsConnected tools
Memory, defended
Continuity was the first step. Integrity is the standard.
HOM is in active development. Join the release list for product and security updates.