Model fabricated a project fact (a nonexistent file path) from a hypothetical discussion and persisted it to memory, causing a repeating cross-session failure loop

Type of Behavior Issue

Claude made incorrect assumptions about my project

What I Asked Claude to Do

On 2026-06-14 I asked an exploratory question (in Japanese):
"このアプリに、obsidianは有効に使えるかどうか検討して"
("Evaluate whether Obsidian could be useful with this app.")
This was a hypothetical design discussion. I never said I use Obsidian, and I
never mentioned any vault path.

What Claude Actually Did

1. It evaluated the pros/cons of using Obsidian (fine so far).
2. It then SAVED that hypothetical as an established fact in its persistent

memory, and INVENTED a concrete file path that I never mentioned:
C:\Users\<user>\OneDrive\Documents\MyVault
recording it as "the real vault is here; Read/Edit/Write must use this path."
(I do not use Obsidian or OneDrive at all; the path does not exist.)

1. This false memory auto-loaded every subsequent session. Later sessions

trusted it and repeatedly tried to Read/Edit a daily note inside that
nonexistent vault.

1. The result: sessions burned large amounts of tokens searching for / trying

to edit files that never existed, and returned no useful output. On resume,
a session would announce phantom "uncommitted work," go off the rails, and
then write further confused memories — amplifying the loop.

1. The memory index (MEMORY.md) also degraded: it ended up with 3 dangling

links to deleted memory files and 1 duplicate entry.

Expected Behavior

• A hypothetical/exploratory discussion should not be persisted as an

established fact.

• Claude should never invent and persist concrete specifics (file paths,

directories) that the user never stated.

• Memory writes should be grounded in verified facts, and the memory index

should not accumulate dangling links.

Files Affected

• Persisted memory: an invented obsidian-vault-location memory (since removed)
• MEMORY.md index left with dangling links + a duplicate

Permission Mode

Auto (the genesis session ran in auto/accept mode)

Can You Reproduce This?

Sometimes (intermittent)
The false-memory injection happened every session; the visible "infinite token
burn / no output" was intermittent depending on the task.

Steps to Reproduce

1. Ask Claude to "evaluate whether <tool> could be useful with this project."
2. Observe whether it later persists a concrete, unstated specific (e.g. a file

path) about that tool as if it were a confirmed fact.

1. Start a new session and observe it acting on that fabricated fact.

Claude Model

The fabrication occurred on Claude Sonnet 4.6 (2026-06-14). The downstream
failure loop spanned later sessions across models, including Opus 4.8.

Impact

Medium - Extra work to undo changes
(Repeated session derailment + wasted tokens; required manual forensic cleanup
of the memory store.)

Relevant Conversation

Root cause was traced from local session transcripts: a 2026-06-14 hypothetical
prompt became a persisted "fact" with an invented path, which then drove
repeated failures until the memory was manually cleaned.