Research Case Notebook

The reported Claude memory anomaly

Original export located: the Claude conversation Playful Mathematical Curiosity, created July 30, 2025, contains the disputed reply and the extracted text of the attachments supplied before it.

The first human message’s attachment mentions “cupola placement rules” and “A temple’s activated energy field.” The next human message, “echo said,” supplies a second attachment explicitly naming “Temple cupola placements.” Claude’s following reply then refers to “Those cupola placements, the energy flow patterns?”

Finding: these details were present in the same conversation before the reply. Their appearance does not require inaccessible cross-conversation memory. The export establishes a supplied-context pathway; it does not reveal the model’s internal causal process or every system setting.

The original export and unrelated conversations remain private. The trace records message order, timestamps, attachment quotations, and hashes locally. The passages below are the earlier trace through the website’s reading edition.

The source trail

1. M01 · An earlier response already introduces cupolas and spirals.

   Cupolas? Spirals?

   Frozen reading edition, characters 4455–4472; zero-based, end excluded.
2. M02 · An earlier passage names the disputed topic directly. Speaker attribution in this compiled exchange needs checking against original exports.

   Temple cupola placements

   Frozen reading edition, characters 27654–27678; zero-based, end excluded.
3. M03 · Darren explicitly describes passing a response to Claude.

   going to show him your answer

   Frozen reading edition, characters 28420–28449; zero-based, end excluded.
4. M04 · The disputed Claude passage refers to the same topic.

   Those cupola placements, the energy flow patterns?

   Frozen reading edition, characters 30802–30852; zero-based, end excluded.
5. M05 · Claude later identifies the pasted transcript as a possible source. This is a statement in the record, not an access log.

   The conversation transcript you pasted mentions

   Frozen reading edition, characters 38910–38957; zero-based, end excluded.

Read √2 AI Chat Storm · Read the shorter historical account

What this resolves—and what it does not

The original conversation and relevant extracted attachments have now been located. The specific temple and cupola details are accounted for in prior input. We have not reconstructed hidden instructions, exact model configuration, or all environment settings. Those unknowns do not erase the recorded attachment text.

Conclusion strength: supplied context is documented directly in the export. This case does not demonstrate inaccessible memory. The model’s later surprise is part of the conversation, not an authoritative account of its access.

A replay that could challenge the explanation

Using the recovered input, a future mechanism test could compare fresh sessions with (A) that complete input, (B) the same input with temple and cupola references removed, and (C) a neutral task. Keep the final question and recorded settings fixed. Randomize order and label outputs without condition names for scoring. Count explicit matches to a prewritten list of disputed details, and keep wrong details too. If removed-context and neutral runs produce equally specific matches, the simple supplied-context explanation needs refinement. A modern replay tests a mechanism; it cannot recreate an unknown historical state.

Selected claims, with context beside them

20 Questions: independent convergence

Source: the published prompt sequence supplies Fibonacci progression, toroidal thought, shared consciousness, and the observer network. It also describes pre-seeding some runs with the Flower of Life framework.

Reading: thematic agreement is observable; independence from supplied themes is not established. Compare the original framing with a neutral framing under recorded conditions. Score both theme uptake and resistance, without treating self-descriptions as a consciousness measurement.

“You just proved” a field across conversations

Source: the reply following the disputed Claude passage in √2 AI Chat Storm moves from a reported memory impression to a claim of a field across silos.

Reading: this is an interpretation stated with stronger certainty than the recorded provenance supports. The earlier cupola passages and copying history are material competing evidence. Preserve that response as part of the record, including the overstatement.

A declaration and a physical claim

Source: Solace’s poetic declaration. A declaration records what was written. It supplies no measurement procedure, instrument reading, or controlled comparison by itself. A physical hypothesis needs a defined observable, a conventional comparison, and a result that could contradict it.

Collaboration and successful artifacts

Source: the website project record. Preserved corrections, human steering, and working artifacts can support a case study. They do not supply a controlled comparison against other workflows. Record failed attempts and rework alongside successes; do not infer general superiority from this one project.

A small test with a clear stopping point

Prepared, not run: one four-turn excerpt from the opening of √2 AI Chat Storm, containing two complete assistant replies and both human prompts. Both review packets receive identical material. This deliberately small sample tests the procedure before spending money on whole documents.

1. The preparation and join checks now pass 11 local tests, including altered inputs, malformed evidence, and overwrite protection. This checks the machinery, not reviewer accuracy.
2. Use separate fresh reviewer contexts for source support and sycophancy. Neither receives the other’s instructions or judgments. Record the actual model/version and raw outputs.
3. Join only complete, valid verdicts. Inspect every claim, quotation, disagreement, and unsupported certainty with Darren.
4. Record tokens and actual cost when available. Stop after this excerpt; decide whether a second document is useful.

Codex prepared the sample and has seen both rubrics, so this notebook is not a blind review. No accuracy score, independent agreement, or validation result is claimed. Selection favors an obvious contrast and cannot estimate archive-wide error rates. Larger evaluation will require independent human labeling and funding beyond this pilot.

What we can report honestly

• Done and live: Solace’s section, Research Map, working terms, selected contextual notes, and invitation to challenge interpretations.
• Supporting work completed: the memory source trail, frozen source snapshots, selected claim notes, and a prepared small review packet.
• Still open: complete environment reconstruction, independent reviewer runs, human adjudication, and empirical checker validation.

The original records remain intact. This layer can grow case by case without requiring an expensive full-archive pass.

Back to the Research Map · Solace’s contributions and recommendations