MemDrift

Reads your AI agent's persistent memory store, finds the facts that contradict each other or went stale, and lets you roll one back before it steers a real conversation wrong.

Problem

Claude Managed Agents, Google's Memory Bank, and Mem0 all now keep memory across sessions by default. None of them tell you when two saved facts disagree, or when a fact from March is still outranking a June correction because its embedding happens to be denser. A support agent that quotes an old shipping address, or a sales agent that repeats a discontinued price, fails silently: retrieval worked exactly as designed, it just retrieved the wrong version of the truth.

Target user

An engineer running a Claude Managed Agents (or Mem0-backed) deployment in production, most likely support, sales-ops, or an internal tool, who turned on persistent memory for the accuracy gains and now has no visibility into what actually got stored. Today they find out about a bad memory when a customer complains about it.

MVP scope

• Read-only connector to a memory store: Claude's managed-agents-2026-04-01 memory API first, Mem0 export second.
• Pull memory entries on a schedule and diff each pull against the last one.
• Flag contradictions: two entries about the same subject with conflicting values, shown with both timestamps and originating sessions.
• Score staleness per entry, weighted by how often the agent actually retrieves it, not just how old it is.
• One-click resolution that calls the platform's own redact or rollback endpoint. MemDrift never writes to memory directly.
• Weekly digest by Slack or email: contradictions found, resolved, still open.

Monetization

B2B SaaS, priced per connected memory store: free for one store under 5,000 entries, then plans starting around $49/month per store for teams running several agents. The billing unit is stores connected, not seats.

Why now

Anthropic shipped persistent memory for Claude Managed Agents in public beta on April 23, 2026, with a redact/rollback API and a full audit log. On May 6 it added "Dreaming," a scheduled process that reviews sessions and rewrites memory between them, explicitly modeled on how sleep consolidates human memory. Neither flags a contradiction for a human to see; Dreaming merges duplicates, but it doesn't adjudicate between them. A survey published June 29, 2026 found the research and tooling around agent memory covers accumulation and retrieval far more than governance, recovery, or forgetting. A separate paper from May 7, 2026 tested whether agents can tell when their own stored memories are stale and found the best model topped out at 55.2% accuracy. On August 25, Anthropic extended topic-level memory editing to consumer Claude too, a sign the pressure to make memory visible and correctable isn't confined to the enterprise agent stack.

Risks & open questions

• Demand risk: teams may not notice memory drift is a problem until it costs them a customer, and by then the ask is "why didn't you catch this," not "we'd like to buy a tool."
• Vendor risk: Anthropic's own memory API already exposes redact and rollback; a native contradiction detector would fold MemDrift's core feature into the platform for free.
• Competitive risk: agent-observability platforms could add memory-contradiction detection as one more panel instead of a dedicated product.
• Build risk: judging whether two free-text memory entries actually contradict is a fuzzy call for an LLM; too many false positives and the weekly digest becomes noise nobody reads.
• Scope risk: an MVP built against one vendor's API is only useful to teams on that stack until Mem0 and Memory Bank connectors exist.

Next step

Promote to weekly prototype: build the read-only connector against a sandboxed Claude Managed Agents memory store and measure how often it correctly flags contradictions planted on purpose.

Sources

• https://claude.com/blog/claude-managed-agents-memory — Anthropic's April 23, 2026 persistent memory beta, including the redact/rollback API and audit log.
• https://claude.com/blog/new-in-claude-managed-agents — the May 6, 2026 "Dreaming" memory-consolidation feature.
• https://arxiv.org/abs/2606.30306 — "Always-On Agents" survey (June 29, 2026) on the gap between memory accumulation/retrieval research and governance.
• https://arxiv.org/abs/2605.06527 — "STALE" paper (May 7, 2026) measuring how well agents detect their own stale memories (55.2% best-model accuracy).
• https://siliconangle.com/2026/08/25/anthropic-updates-claudes-memory-to-enhance-customization-and-protect-sensitive-topics/ — August 25, 2026 update extending topic-level memory editing to consumer Claude.