Radar / Ideas / MemoryLedger: tamper-evident audit of…

MemoryLedger: tamper-evident audit of what your agents remembered and decided

daily ideaambitiousJEV confidence 0.62JEV monetization 0.612026-10-18
OutcomeRegulated companies deploying AI agents pass EU AI Act Article 12 audits with a tamper-evident, replayable record of every agent decision and the exact memory state behind it.

The problem

EU AI Act Article 12 requires automatic, tamper-evident logging of high-risk AI decisions, retained at least six months, with penalties up to €15M or 3% of global turnover. Most deployments produce zero compliant records. Worse, existing agent-governance tools log actions — not the memory state the decision rested on — so when an agent hallucinates from poisoned or stale memory, the audit trail cannot explain why.

The idea

A memory-layer sidecar (integrating with Mem0/MemOS/Hy-Memory-style stacks) that snapshots memory state at every agent decision, chains it with inputs, outputs, and approvals in a tamper-evident signed log, and produces an Article-12-ready audit pack: what the agent knew, what it decided, who approved, when. Includes a replay mode that reconstructs any past decision with its exact memory context — the question auditors ask after a hallucination incident, which action logs cannot answer.

Why now

Hy Memory 2.0 (Tencent, Oct 16) made versioned, cross-tool agent memory production-grade; decision models (d1-ultra, StartLux v1.1) made cheap calibrated judgments available; and the EU AI Act's high-risk obligations took effect on 2 August 2026 — the compliance deadline is now, while DeepInspect reports most deployments produce zero compliant records.

What it combines

Versioned agent memory (entry tencent-hy-memory-2-0-production-grade-dual-system-cognitive-memory-engine) crossed with typed decision judgments (entry jev-1-14-0-typesafe-s-judgment-layer-update): memory snapshots plus typed verdict records produce a replayable, auditable history of the agent's beliefs — not just its actions.

MVP

Weekend MVP: a wrapper around a Mem0-compatible memory client that (1) snapshots memory before/after each agent decision, (2) appends the decision plus a memory hash to an append-only signed log, (3) generates a one-page Article-12 audit summary for a demo incident. Deliberately skip: cross-tool integrations, replay UI, retention policies, SIEM export.

How it makes money

Budget owner: CISO / Head of GRC — already pays for compliance automation. Pricing anchor: Drata $10k–$80k/yr and Vanta $10k–$100k/yr (buyer-reported pricing), with audits costing $10–30k extra. Willingness-to-pay evidence: Drata launched AI Agent Governance in June 2026 because buyers are already paying for agent governance; EU AI Act Article 99 sets penalties at €15M or 3% of global turnover; 89% of companies leave agent-governance questions unanswered. Replacement budget: compliance spend already exists; this fills the memory-evidence gap in it.

The moat

Drata and Vanta are architected around policy-and-inventory governance (agent discovery, config evidence, human attestation). Memory-state capture requires instrumentation at the agent memory layer — a different architecture they'd have to build from scratch — and selling 'your action logs aren't enough for incident forensics' would undercut their own evidence story. Data moat: the corpus of memory-diff crossed with incident pairs across deployments, which action-log vendors never collect.

Distribution

Sell to CISOs and Heads of GRC at banks, insurers, and healthcare companies deploying agents; wedge on EU AI Act compliance deadlines; partner with compliance platforms (Drata, Vanta) as the agent-memory evidence source they don't have.

Why it wins

Drata's AI Agent Governance (June 2026) inventories agents and logs actions in tamper-evident records; Vanta and Drata sell config-state evidence. None capture memory state — the layer where hallucinations and prompt-injection persistence actually live. This audits the agent's beliefs, which is a different evidence dimension from action logs.

Risks

Buyers may accept action-level logs (Drata) as 'good enough' — de-risk by piloting with one compliance team and asking their auditor directly whether memory-state evidence closes a gap in their Article 12 pack. Second risk: snapshot overhead; de-risk with hash-only snapshots plus on-demand full capture.

Build it with

Repo to start from

memory-ledger — tamper-evident audit trails for agent memory: snapshot what the agent knew at every decision, replay any incident.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.