Keon Cortex / Deterministic Memory

One command. Seven proof families. Content-addressed evidence.

Keon Cortex is a deterministic AI memory substrate with a unified proof bundle, authoritative truth separation, benchmark artifacts, and governed Evidence Pack verification.

Purpose

Define the memory substrate responsible for canonical truth, receipt-backed reconstruction, and proof-oriented verification.

Primary question

How do you verify that the memory layer is correct?

Allowed
  • unified proof bundle
  • I1-I7 proof spine
  • content-addressed proof identity
  • Evidence Pack verification
Forbidden
  • vector database positioning
  • chat history cache framing
  • runtime execution governance
  • canonical truth stored in the index
What It Does

Proof-oriented capabilities.

  • Separates canonical memory truth from disposable retrieval surfaces. The vector index is derivative. The document store is authoritative.
  • Publishes cortex proof-bundle as a verify-all artifact that composes I1-I7 into one portable proof bundle.
  • Makes the proof bundle content-addressed through a deterministic Merkle root where proof_bundle_id equals merkle_root.
  • Emits stronger temporal memory receipts and retrieval influence receipts so buyers can inspect what memory existed and why it influenced retrieval.
  • Generates deterministic proof and benchmark artifacts, with governed mode adding signed receipts and tamper-detectable Evidence Pack verification.
What It Is Not

Boundary protected.

  • Cortex is not a vector database, RAG wrapper, or chat history cache.
  • Cortex does not decide whether an action may execute.
  • Cortex does not perform runtime execution governance or turn retrieval output into authority.
Boundary Definition

Memory is evidence. Retrieval is disposable. Permission lives elsewhere.

Cortex records, reconstructs, and verifies canonical memory while Runtime remains the execution boundary and Collective remains the cognition layer. Vector databases are semantic indexes. They are not memory.

System Connections

Receives, outputs, never.

Receives from
Runtime
receipt-backed events and execution lineage
MCP Gateway
durable ingress receipts when gateway persistence is enabled
Control
inspection requests for lineage and proof surfaces
Outputs to
Control
temporal receipts, influence receipts, proof bundles, manifests, and replayable state
Proof
content-addressed proof bundles, invariant manifests, benchmark artifacts, and verification material
Collective
authorized context, never execution permission
Never does
decidesexecutesmutates history
Verifier-first positioning

Memory claims require proof.

Vector databases are not memory. They are semantic indexes.

Cortex separates canonical memory truth from disposable retrieval surfaces.

One command. Seven proof families. Deterministic output. Content-addressed evidence.

Do not trust the memory layer. Verify it.

Unified proof bundle

One command. Seven proof families. Deterministic output.

Command
cortex proof-bundle [--output FILE]

Single verify-all export for the full Cortex proof spine.

Coverage
I1-I7

Composes deterministic memory, tenant isolation, replay durability, authority/index separation, shard identity, decay/reinforcement, and governed signing evidence.

Identity
proof_bundle_id == merkle_root

The bundle is content-addressed by deterministic SHA-256 Merkle root.

Gate
byte-identical double-run

CI can diff repeated outputs and reject non-deterministic proof generation.

The proof bundle is the courtroom binder: portable, inspectable, CI-gated, and ready for downstream Evidence Pack integration.

Invariant manifest

Proof spine I1-I7.

I1cortex proof
Deterministic memory / core proof summary

Temporal memory receipts and the core proof summary are deterministic and byte-identical across proof gates.

I2cortex tenant-isolation-proof
Tenant isolation

Fail-closed tenant isolation rejects missing, empty, aliased, and foreign tenant paths before memory can leak.

I3cortex replay-proof
Replay / outbox durability

Replay, leasing, restart, retry, and outbox durability prove no lost writes across deterministic reconstruction paths.

I4cortex authority-proof
Authoritative store / derivative index

Canonical truth lives in the authoritative store while the vector index remains disposable and replayable.

I5cortex shard-identity-proof
Deterministic shard identity / idempotent ingestion

Identical canonical payloads resolve to the same shard identity and idempotent retries do not create duplicate records.

I6cortex decay-reinforcement-proof
Decay & reinforcement

Trust decay and reinforcement remain deterministic, with floors and ceilings preventing poisoned memory from winning by mass alone.

I7cortex governed-signing-proof
Governed signing / Evidence Pack integrity

Governed mode signs receipts with Ed25519 and verifies tamper-detectable Evidence Packs without making signing mandatory for open-core receipts.

Recent hardening delta

Receipts and memory semantics got stronger.

Temporal receipts

Stronger temporal memory receipts make the evolution of memory state reconstructable, not merely asserted.

Influence receipts

Retrieval influence receipts show selected, excluded, ranked, and degraded memory paths instead of hiding retrieval behavior.

Fact identity

Fact slots, duplicates, and reinforcement remain distinguishable so strengthening a claim does not collapse lineage semantics.

Durable replay

Outbox leasing, safe reprocessing, restart tolerance, and idempotent replay behavior are proof-backed rather than implied.

Trust hardening

Decay and reinforcement are hardened so low-trust memory cannot become authoritative by repetition alone.

Deterministic artifacts

Proof and benchmark outputs are generated as stable artifacts suitable for CI upload, comparison, and independent review.

Proof and benchmark artifacts

Deterministic outputs you can audit.

Unified proof bundle
cortex proof-bundle

One portable auditor-grade bundle composing I1-I7 with content-addressed Merkle identity and all_checks_passed aggregate status.

Proof artifacts
cortex proof • cortex invariant-manifest

Core proof summary, invariant manifest I1-I7, deterministic JSON output, and byte-identical proof gates.

Benchmark artifacts
cortex benchmark

Deterministic benchmark reports with provider metadata, stable ordering, CI upload support, and regression-grade comparison outputs.

Governed Evidence Packs
governed mode verification

Ed25519 receipt signing, deterministic manifest/hash tree generation, tamper detection, and historical-key verification coverage.

Open core

Unsigned open-core receipts remain allowed where governed signing is not enforced.

Governed mode

Governed signing, key lifecycle support, and Evidence Pack integrity become the commercial differentiator for verifier-heavy deployments.

CI now uploads proof and benchmark artifacts, invariant-manifest checks cover I1-I7, proof_bundle_id equals the deterministic Merkle root, and byte-identical double-run gates keep proof outputs reproducible across reruns.

Why Cortex differs

Remembering useful context is not enough.

Mem0, LangGraph/LangMem, Letta, LlamaIndex, Zep/Graphiti, Cognee, and Hindsight mostly optimize for agents remembering useful context. Cortex optimizes for proving memory correctness under failure, tenancy, replay, indexing drift, trust decay, and governed audit.

Verdict: Cortex leads on durability/replay proof. Transactional outbox, leasing, replay proofs, idempotence, and byte-identical artifacts are not the normal pitch in this market.

Market contrast

Most memory systems optimize recall. Cortex optimizes correctness.

Recallcorrectness

Most memory systems optimize recall. Cortex optimizes correctness.

Context retrievalmemory lineage

Most memory systems retrieve context. Cortex proves memory lineage.

Index trustindex disposal

Most memory systems trust their indexes. Cortex treats indexes as disposable derivatives.

After-the-fact logsdeterministic proof artifacts

Most memory systems log after the fact. Cortex emits deterministic proof artifacts you can verify.

Cryptographic auditability

The commercial moat is audit infrastructure.

Ed25519 signing

Governed mode signs receipts with Ed25519.

Valid signature acceptance

Verifier accepts valid signatures.

Tamper rejection

Verifier rejects tampered receipt payloads.

Revoked key rejection

Verifier rejects signatures from revoked keys.

Retired key acceptance

Verifier accepts signatures from retired historical keys.

Deterministic Evidence Pack

Manifest/hash tree is deterministic and tamper-detectable.

Proof bundle composition

Governed signing composes into the unified proof bundle as I7.

Verdict: Cortex leads on cryptographic auditability. Ed25519 signing, tamper rejection, revoked key rejection, retired key acceptance, deterministic Evidence Pack manifest/hash tree, and proof bundle composition are not memory features. They are audit infrastructure.

Next Action

Continue through the correct surface.

Verify a receipt