BY NEURABYTE
— THE PLATFORM

What your agents knew, were allowed to do, and actually did — on one record.

Most tools hold one part of the problem: memory, traces, content inspection, or after-the-fact evidence. Lore holds the whole chain — and every memory mutation, compiled context block, and governed action leaves a signed, tamper-evident receipt on the same record.

/01 · GOVERNED MEMORY

Shared institutional knowledge agents can act from.

Memory follows the organization and the project — not an individual developer, IDE, or model. Curated snapshots give an agent the smallest reliable body of knowledge for the task.

Enterprise Bootstrap
One set of instructions for every agent, layered from system and organization down to team, project and user, delivered at session start.
Memory Studio
Block schemes, per-client bindings, and snapshots from a compiled block.
Bi-temporal facts
Valid time and transaction time on every fact; contradictions close records instead of deleting them. "What did the agent believe on June 3?" is an ordinary query.
Hybrid retrieval, benchmarked
BM25 + vector + graph retrieval fused — 0.98 recall@8 on our internal benchmark suite.
Review before it shapes behavior
Approval inboxes, decay policies, pin and archive states, sensitivity flags, trust classification.
Provable deletion
Cryptographic erasure with per-unit key destruction — a receipted, offline-verifiable record of exactly what was destroyed.
Counterfactual replay
Re-run a task with selected memory withheld, non-actuating. A signed influence record answers whether a bad memory caused a bad action.
Capture at the source
Six verified capture clients — Claude Code, Codex, Cursor and more — at 1.86 s p95, stored as typed transcript parts.
CLAUDE CODE CODEX CURSOR capture · 1.86 s p95 01 REVIEW maker–checker approve BI-TEMPORAL RECORD t1 t2 t3 as-of t2 query recall@8 · 0.98 CURATED SNAPSHOT receipt
/02 · CONTEXT MANAGEMENT

Control exactly what an LLM is allowed to see.

Possessing information doesn't mean it should be sent to a model. Lore governs context by organization, tenant, project, user, agent, task, and model destination — and proves what was actually transmitted.

Context Materialization
For repeated workloads, context is prepared ahead of time with versioned recipes, preparation jobs and cache residency, for hosted providers or your own vLLM and SGLang.
Six states, not one permission
Possessed → searchable → available → approved for a model → selected for the task → actually transmitted. Each state is distinct and governed.
Policy at the prompt boundary
Include, exclude, redact, summarize, transform, or require approval before content enters a prompt.
Compiled blocks with provenance
Every assembly produces a canonical block with a stable SHA-256 identity and Ed25519 receipt, stamped onto the execution it fed.
Immutable context snapshots
Source artifacts and versions, retrieved facts, transformations, exclusions, and the final payload — preserved exactly as the model received it.
Progressive disclosure on MCP
13 tools up front, 88 more discovered on demand, served over HTTP. Cheap enough to carry into every session.
Analytical history at scale
An Apache Iceberg lakehouse with point-in-time as-of reads: which files were disclosed to which model, at any moment.
COMMIT PROVENANCE

Every commit knows which agent sessions produced it.

Lore links sessions to commits as they happen, with a Lore-Session trailer on agent commits. The Commits view shows linked sessions, the diff and the evidence together, and the code map traces a change from repository to commit, file and symbol, with blast radius before it merges.

CONVERSATIONS · COMMITS
Commits view linking repository commits to the agent sessions behind them
repository
commit
file
symbol
Code map: trace a change down to the symbol, with blast radius before merge.
/03 · LORE STATE BRIDGE · VALUE PROPOSITION →

Governed execution for agents operating real applications.

Agents report their own success. Lore doesn't take their word for it: it authorizes before, verifies after, and signs the result. We have caught a lying client in live operation.

authorize → approve → execute → verify → receipt
Approvals where people already are
Approvals in Slack and Microsoft Teams, and tickets in Jira and ServiceNow, configured per project. The decision still runs through the gate.
A blocking gate on every surface
In-app chat, agent socket, MCP tools, and REST all use the same authorize–execute–complete cycle. Fail-closed behavior is regression-tested.
Maker–checker approval
Destructive operations need scope-bound, single-use approval tokens from a distinct authorized checker.
Independent post-state verification
The server re-derives authoritative state after execution. Contradictions are recorded as state_mismatch, never silently accepted.
Signed receipts, every outcome
Verified, attested, mismatch, failed, and blocked outcomes all produce canonical receipts dual-signed with HMAC-SHA256 and Ed25519.
Fleet operations
Token budgets by agent, project, or org; kill switches; heartbeat failover; anomaly detection; signed skill manifests.
A protocol, not a feature
@lore/contracts, an OpenAPI surface, and SDKs in TypeScript, Python, Go, Rust, .NET, and Java — other frameworks can adopt the gate without adopting our UI.
AGENT LORE APPLICATION authorize(op) policy check maker–checker scoped token · single-use execute verify post-state authoritative fields signed receipt no token → no execution · fail-closed
/04 · COMPLIANCE & GOVERNANCE

Enforcement data as an audit product.

Authorization gates, signed executions, and policy decisions already exist because the platform needs them to operate safely. Compliance uses that operating evidence — instead of reconstructing months of activity with screenshots and tickets.

Assurance programs
SOC 2 and ISO 27001 programs with readiness packages, bring-your-own frameworks, Lore identity data as evidence, and the assurance workbench.
Enforcement-derived controls
A 15-control registry mapped across SOC 2 CC1–CC9, evaluated against evidence from the running platform. Absent evidence is recorded as a gap, never a pass.
Signed evidence bundles
Canonical bundles, hashed, Ed25519-signed, hash-chained into an insert-only store. Auditors verify offline: VALID, TAMPERED, or INVALID.
Frameworks as data
SOC 2, HIPAA, GDPR, DORA, the EU AI Act and future regimes normalize onto one canonical model, with versioned crosswalks.
Separated write authority
A dedicated Go runtime owns assurance state (~900 tables). An application bug cannot silently rewrite the evidence layer.
PLATFORM gates · policies evidence 15 CONTROLS SOC 2 CC1–CC9 bundle SOC 2 · GDPR · DORA · EU AI ACT hashed · signed hash hash OFFLINE VERIFIER no database access VALID TAMPERED INVALID absent evidence = gap, never a pass
/05 · TRUST SUBSTRATE

Native PKI, KMS, SSH, and PAM.

Evidence is only as strong as the keys that sign it. Lore ships its own security runtime in Go, so signing, key management, and privileged access live under the same governed record they protect.

Secrets that go where code runs
Secret sync and rotation, certificate sync to AWS Certificate Manager and Azure Key Vault, and secret scanning.
Privileged access
An SSH certificate authority and just-in-time privileged access, under the same governed record.
Keys under governance
Certificate issuance and key operations produce the same receipted, auditable record as every other governed action.
Privileged access for agents
SSH and PAM services designed for a world where the privileged operator is often an agent, not a person.
Enterprise custody options
BYOK and HSM custody with managed receipt-key rotation on the enterprise tier.
ROOT OF TRUST PKI KMS SSH PAM certs keys access sessions hash(n) hash(n+1) rotate keys GOVERNED RECORD BYOK · HSM CUSTODY
/06 · DOCUMENT CUSTODY

Governed custody for what agents read and produce.

Documents are context sources and action outputs at once. Lore holds them in fail-closed custody, so what an agent read — and what it wrote — is part of the same provable record.

Fail-closed by default
Access without authority doesn't degrade gracefully — it doesn't happen.
Custody tied to context
Document versions appear in context snapshots, so disclosure to a model is a recorded, attributable event.
Selective disclosure
Reveal a redacted portion of a receipt chain to an auditor or counterparty without breaking chain verification.
FAIL-CLOSED CUSTODY v1 v2 v3 every write versioned AGENT authorized read no authority · blocked MODEL recorded event context snapshot
/07 · PATENT DESK

Invention discovery, inventorship, and disclosure control.

When agents contribute to invention, provenance becomes a legal question. Patent Desk turns the governed development record into evidence of who invented what, when — and what was disclosed to whom.

Invention discovery
Surface patentable mechanisms from the development record itself, as they emerge.
Inventorship from evidence
Reduction-to-practice trails grounded in the same receipted record that governs development.
Proven in anger
The workflow Neurabyte uses for its own patent program.
GOVERNED DEV RECORD DISCOVER patentable signals mechanism record ATTRIBUTE s.chen · claims 1–7 m.ortiz · claims 8–12 disclosure control FILING PACKAGE · RECEIPTED

Every session makes the record stronger.

Every captured session deepens the governed record your agents act from — and the evidence your auditors accept.

Request a demo
Also from Neurabyte: Cupel
Lore's core mechanisms are patent-pending. Product names are trademarks of their respective owners; listing an integration doesn't imply a partnership.
© 2026 NEURABYTE INC.