Memory your agent is actually allowed to use.
Your AI agent acts on whatever its memory returns — including facts it should never have kept. The Agent Memory SDK makes every recall an authority-checked act: out-of-scope memory is withheld fail-closed, every returned fact carries a content hash and signed provenance, and every decision is auditable back to the exact memory that drove it.
The problem
A clean prompt still spends authority on the wrong memory.
Vector recall returns whatever is nearest — not whatever the agent was permitted to remember. That gap is where agents leak, over-reach, and become impossible to audit after the fact.
Out-of-scope recall
A support agent pulls a customer's payment history into a task that never had permission to see it. The model can't tell — the memory just showed up in context.
Stale & poisoned facts
An expired preference or an injected "remember that…" drives a live decision. Nothing checks whether the fact is still valid, or whether it was ever trustworthy.
No trail back
When a decision is challenged, you can't prove which memory caused it — so every incident becomes a manual reconstruction weeks later, if at all.
The solution
Recall becomes a governed action, not a free lookup.
The SDK sits between your agent and its memory store. Every read is resolved against purpose and scope before a single fact reaches the model — and everything that does reach it arrives with proof.
Authority-checked recall
Each read resolves against purpose + scope at the moment of recall. Out-of-scope memory is withheld — fail-closed, never "best effort".
Content-hashed provenance
Every stored and recalled item carries a content hash and signed provenance, so an auditor can prove exactly which memory drove a decision.
Quarantine & revocation
Suspect or superseded memory can be quarantined or revoked; the engine emits the governance event family on every state change.
Why it pays off
Turn "trust me" into evidence you can hand a regulator.
The returns are capability, not slideware: fewer incidents, faster audits, and a memory layer that fails safe instead of failing silent.
For engineering
Drop-in client in front of your existing store. Recall returns scoped, hashed items — you keep your vector DB, you gain the authority boundary. Fail-closed means a misconfigured scope blocks a leak instead of shipping one.
For risk & compliance
Every agent decision is reconstructable from signed evidence, so an audit is a query, not a project. Quarantine and revocation give you a real off-switch for memory that shouldn't have been kept.
Get started
Governed recall in three lines.
Backed by the KYE Memory Engine™. The wire contract and verification surface are open — verify a decision's memory from published keys alone, with no service in the loop.
# TypeScript / JavaScript npm install @kye/agent-memory # Python pip install kye-agent-memory
import { MemoryClient } from "@kye/agent-memory"; const mem = new MemoryClient({ trustDomain: "acme" }); // Recall is authority-checked at the moment of read const r = await mem.recall({ agent: "support-bot", purpose: "resolve-ticket", query: "customer payment history", }); // r.items each carry contentHash + signed provenance; // out-of-scope memory is withheld, not returned
Give your agent a memory it can defend in an audit.
Authority-checked recall, content-hashed provenance, fail-closed by default. Apache-2.0, no runtime lock-in.