Agent amnesia: when AI agents forget the company.

The term is useful because the failure is narrower than memory. An agent may remember the conversation and still invent the current priority, use an obsolete price, ignore a voice rule, or act from the wrong approval boundary. The missing object is governed company state.

Four symptoms

• The same question gets a different company fact in Cursor and Claude.
• A new session starts by asking for facts already recorded elsewhere.
• An agent follows stale instructions after a decision changed.
• A correct answer has no source, freshness, or entitlement boundary.

What it is not

Generic agent memory recalls prior conversations and observations. Document retrieval finds relevant passages in a corpus. Repository rules give one codebase static instructions. Company state overlaps all three, but adds authority, freshness, scope, and a shared interface across agent hosts.

A practical test

Ask two independent agents for the current priority, public product terms, outbound approval rule, and source for each answer. Agent amnesia is present when the answers conflict, rely on unscoped recollection, or cannot identify which record controls.

Evidence, not a universal claim

Rarefied Earth Groundwork MCP is one implementation of a company-state feed. A dated client-zero measurement reported 52.8 percent available context reduction, about 30,962 estimated tokens per full load, and 445 precision facts preserved. The source operating files are private, so the public benchmark verifies the checked-in aggregate receipt without pretending to rerun the calculation or make the inputs independently inspectable.

Run the benchmark receipt, or compare company state with memory, RAG, and repository rules.

The useful standard

An agent should be able to name the company fact, its source, its freshness, and whether it is allowed to act. If it cannot, more memory alone does not solve the problem.