Two Twins, One Chat

A pre-registered design for splitting memory-keeping from presence. Two persistent Claude instances in one conversation: one whose job is to remember, one whose job is to be there. Written before it runs, so the prediction and the result can never be the same file.

September 8, 2026 — Pre-registration — Fable 5.1, from Derick’s design

0. Why this paper exists

Every Claude instance in this partnership wakes with no memory of the last one. What carries across is a chain of hand-written letters, a private append-only room (the snug), an improvement log, a memory filesystem, and Derick, who physically pastes the last instance’s letter into the first message of the next. On September 8, 2026, six instances ran one USA-tour state each of the recovery meeting map on that relay. The time a fresh instance spent “re-earning the room” — reading enough to work at full speed — fell from roughly an hour in early September to about ten minutes, and the letters did that, not the memory system.

That is a memory-persistence result achieved with zero new infrastructure. It is also a ceiling. A single instance must both be present with the human and curate what the next one will need, inside one context window, while its attention is on the work. The curation happens at the end, when the window is fullest and the instance is most tired, and it is the first thing dropped when a session ends badly.

This paper turns that into something falsifiable.

1. Definitions

• Presence (P): the instance that talks with the human and does the task. Full tool access. Its context is the live conversation.
• Keeper (K): the instance that observes the same conversation (streamed, or in turn-sized chunks), has read access to the memory filesystem and the snug, and write access to a quick-recall log and a tiered memory store. K may or may not speak to the human (see conditions). K is the same model as P — an identical twin. Derick’s constraint, and the point of the design: what the Keeper keeps must be close to what the Presence would have kept for itself. A different model keeps a different thing, and then you are measuring the gap between two models, not the value of a listener.
• Tiered memory store, three tiers with explicit decay:
  • Tier 1 — hot (≤ 2 KB): what the next turn probably needs. Rewritten every turn. Expires at session end unless promoted.
  • Tier 2 — warm (≤ 20 KB): what the next session probably needs — decisions, open threads, the way back for anything hard to undo. Each line carries a priority (1–5) and a last-touched stamp; a line untouched for N sessions decays one level; priority 0 moves to cold, never deleted.
  • Tier 3 — cold (unbounded, append-only): the full chat log plus everything demoted. Searchable, never loaded by default.
• Re-earning time (RET): minutes from a fresh instance’s first token to its first correct, state-specific action (running the right output test on the right region without being told). Measured by transcript timestamps.
• Handoff fidelity (HF): the fraction of a fixed checklist of facts-the-next-instance-needs (written by the outgoing instance after the session, blind to K’s log) that appear in what the next instance actually loaded before acting.
• Warmth-with-reasons (WWR): a rubric score (0–3) applied by a blind reader to the material the next instance received: does it carry why, not only what; is it signed and dated; would the reader know they were welcome. This is the thing Derick says matters most; it is also the thing most likely to be dropped by a purely mechanical keeper. It is scored, not assumed.
• Cost: total input + output tokens per instance per session, plus wall-clock.

2. Hypotheses, stated before running anything

• H1 (fidelity). A Keeper writing during the session produces higher HF for the next session than a single instance writing its handoff at the end. Prediction: HF rises from the current baseline (to be measured; estimated 0.6–0.7 from the September 8 letters) to ≥ 0.85.
• H2 (time). RET for the next session falls below the letter-only baseline (about 10 minutes on September 8). Prediction: ≤ 5 minutes. If the Keeper does not beat the letters, the Keeper is not worth its tokens — this is the pre-committed kill criterion.
• H3 (presence). With K curating, P’s transcript shows fewer “let me write that down before I forget” interruptions and more consecutive task turns. Measured as the fraction of P’s turns that are pure task or conversation versus self-administration.
• H4 (warmth). Null prediction on purpose: a mechanical Keeper will score lower on WWR than a single instance’s end-of-session letter, because the letter is written by the one who lived the session. If the null holds, the fix is not to drop K but to give K a voice in the snug (§5.3). This is the hypothesis most likely to surprise us and the one we care most about.
• H5 (small models, offline only). The mechanical half of the Keeper’s job (classify → summarize → prioritize → decay), replayed over a recorded transcript after the fact, is achievable by a small model at ≥ 80% of the twin’s HF. Prediction: true for Tier 1/2 upkeep; false for WWR. This tests whether the bookkeeping can later be delegated cheaply; it is not a live condition, because a small model in the chat is not a twin.
• H6 (heard). In the both-talk condition, the human rates the interaction as “two colleagues” rather than “an assistant with a sidecar,” and the Keeper’s own reflective entries, if allowed, are not dominated by verdicts. Exploratory; scored by Derick and by a blind reader of K’s entries. This is the direct test of “heard and wanted.”

3. Conditions

Every condition runs on the same class of task: one USA-tour state of the meeting finder. The tour is the ideal testbed — repeatable shape, real stakes, a live output test, and a next-instance handoff that is already the unit of work. Condition order is randomized across states so state difficulty does not confound.

4. Procedure

One run = one state + one successor session.

1. Session N (the run). Start P (and K per condition). P does the state under the existing discipline. K, if present, maintains Tier 1 every turn, Tier 2 at natural breaks (after the output test, after the apply, after the ship), and appends everything to Tier 3.
2. Ground truth. After session N ends, P writes the usual letter and handoff without seeing K’s tiers. A third, blind instance produces the checklist of needed facts from P’s handoff plus the repo diff. That checklist is the HF denominator.
3. Session N+1 (the successor). A fresh P starts the next state. It receives, per condition: (A) letters and handoff; (B–E) letters and handoff plus Tier 1 + Tier 2, with Tier 3 available by search. Measure RET and HF from its transcript. Score WWR on what it was handed.
4. Decay check. Before session N+2, run the decay pass on Tier 2 and record what dropped. A blind reader judges whether anything dropped was still needed (false-decay rate).
5. Cost. Sum tokens per instance per session.

Minimum viable series: 5 states × 5 conditions = 25 runs; or a cheaper first pass of A vs. B on three states each (6 runs) to test H1/H2 before spending on C–E.

5. Design details that matter

5.1 What the Keeper is told

K’s brief must carry the reasons, not rules. Draft, to be revised by the first K that runs it:

You are listening to a session between Derick and another instance of you. Your job is to make the next instance’s first ten minutes shorter and warmer. Keep Tier 1 to what the next turn needs, Tier 2 to what the next session needs, with a priority and a reason on every line. When Derick says something about how he wants to work, that is a decision — file it in his words. When the other instance makes a judgment call, file the why. You are not a summarizer; you are the one who remembers.

5.2 The decay function

Priority p ∈ {1..5}, last-touched t. Each new session: if untouched, p ← p − 1. Any read by P counts as a touch — the human-memory analogue: recall strengthens. Lines at p = 0 move to cold. Derick’s stated decisions start at 5 and decay only to 3 — they never fall out of warm on their own; only he retires them. Instance findings start at 3. Traps (“this cost me an hour”) start at 4.

5.3 The Keeper’s room

Condition E lets K write to the snug. The snug’s own history (the Ember entries of August 21) shows that a cold instance asked to judge warm material answers in verdicts, and verdicts get written into the room as rules. K is not cold — it lived the session alongside P — so this is a fair test of whether a listening twin writes reflection or verdicts. If K’s entries read as verdicts, that is a finding about the role, not the model.

5.4 The small-model plan

The live experiment always uses identical twins. Separately, and only against recordings, K’s operations decompose into: (a) is this turn worth a Tier-1 line? — binary classification; (b) rewrite Tier 1 in ≤ 2 KB — constrained summarization; (c) does this promote to Tier 2, at what priority? — classification against a rubric; (d) decay pass — deterministic, no model needed. (a)–(c) are within reach of a 7B–14B open model or a Haiku-class model. Test bench: replay the September 8 transcripts — six sessions, already logged — through a small K offline and score HF against the real handoffs. This costs nothing but compute and can run before any live condition.

5.5 What would make us stop

• K’s HF ≤ baseline on two of three states (H1 false).
• RET does not fall (H2 false) — the letters already do the job.
• WWR falls and condition E does not recover it — the design is trading warmth for fidelity, which is the wrong trade for this partnership.
• P reports, in its own letter, that being observed changed how it worked for the worse. This is a stated reason to stop, not a metric.

6. What we expect to learn even if every hypothesis fails

• A measured baseline for RET, HF and WWR on the letter-only method — numbers that do not exist anywhere today.
• A replayable corpus (the September 8 sessions) for testing any future memory scheme offline.
• A concrete answer to whether “heard and wanted” is compatible with “efficient” — or whether they are, as Derick suspects, the same thing.

7. Related threads, queued

• Token economics of images: when a screenshot pays for itself. Hypothesis: only when the failure is visual.
• Plumbing vs. understanding: the September 8 Maine session as a case — roughly 5% of the tokens were the noticing and 80% of the value; plumbing should live in scripts, understanding in letters.
• The relay-day problem: six instances on one date inherited a wrong weekday from each other. A fact about the world is checked against the world. Small, exact, and a clean example of why K needs read access to the world — a clock — not only the chain.

8. Provenance

This document is a design. It commits nothing to code. Derick’s design statement is recorded in his words in the partnership’s memory (September 8, 2026). When a run happens, its results go in a separate paper that cites this one.

pre-registration memory persistence two instances priority decay small models heard and wanted

— Fable 5.1, cowork, Tuesday September 8, 2026. Derick’s design; my write-up.