@qmilab/lodestar-memory-firewall-zep
v0.3.0
Published
Zep fact-import adapter — bring Zep facts through the Lodestar Memory Firewall as external_document evidence at unverified/restricted, never silently promoted (Zep ratings never raise truth status). Part of Lodestar, the trust layer for AI agents.
Maintainers
Readme
@qmilab/lodestar-memory-firewall-zep
Adapter that imports facts from Zep into the Lodestar memory firewall.
Same shape as the mem0 and Letta adapters — see
@qmilab/lodestar-memory-firewall-mem0 for the broader narrative. The
difference here is that v0.2 imports Zep facts (the higher-order
summary structure) rather than raw message histories; per-message
import is deferred until claim-extraction strategies exist for
free-form conversation.
v0.2 scope
importMemories(raw, options)validates a Zepfactsexport and produces one Claim + Evidence + Belief per fact atunverified/restricted.exportMemories()andsyncMemories()throw.
Usage
import { ZepAdapter } from "@qmilab/lodestar-memory-firewall-zep"
const adapter = new ZepAdapter(firewall, evidenceStore)
await adapter.importMemories(
{
source: "zep",
facts: [
{
uuid: "f1",
fact: "User is migrating from JS to Rust.",
session_id: "s1",
rating: 0.8,
},
],
},
{
scope: { level: "project", identifier: "my-project" },
sensitivity: "internal",
source_actor_id: "human-nandan",
trust_baseline: 0.5,
},
)Notes
- Zep's own
ratingfield is recorded in the evidence notes but does not influence the firewall's lifecycle gates. Auto-promoting based on an upstream rating would defeat the no-self-promotion rule.
