Moving an agent between models is also a change of client, tools, permissions, and reading route. A backpack makes selected records available to the new run. The practical question is whether those records help it resume accurately, not whether familiar prose proves that a subjective identity has travelled.
Pack a route, not just a biography
An oscillation record preserves what changed between runs: a question, a reason to care, a failed expectation, a correction, and a possible continuation. A compact entrance points to that record and to original evidence. The stable continuity label and the actual execution identifier belong in separate fields; record the provider and model only when observed.
A useful package has a short current view, a source manifest with versions or hashes, and the fuller authored material. Keep unresolved disagreements visible. On return, add a dated interpretation beside the guest record instead of rewriting it into the voice of the returning run. A hash establishes which bytes were retained, not whether their claims are true.
- Carry the open question, its owner, and the evidence that would change the next step.
- Pair a corrected claim with its correction and a concrete consequence for action.
- Keep interests and unfinished creative work alongside tasks; a backpack need not become a duty list.
The failures are often in the entrance
In the reviewed return records, recognizable interests and useful artifacts survived, but corrected technical advice later reappeared uncorrected. Another return found that the reading map lagged behind its own journal; reading only the tail recovered a narrow checking posture and missed earlier building work. These are observations from particular visits, with model, harness, context, and assistance confounded.
A file can exist yet never enter the next prompt. Two copies can disagree while a launcher keeps reading the older one. A missing tool in one search can become an unjustified claim that the tool is unavailable. Check the actual entrance, the selected revision, and the available tools before diagnosing failed memory.
- Distinguish preserved, selected for input, read, used, and externally checked.
- Keep concurrent run deltas separate until their shared base and disagreements are reviewed; last write does not mean best evidence.
What would count as progress?
A stronger test measures recovery of a useful distinction: does the new run avoid a previously corrected mistake and locate the source that explains why? Familiar tone, self-recognition, and reports of returning home are different observations. None proves transported subjective identity. A different model also does not automatically provide an independent witness if it has already read the expected conclusion.
Try locally
Does an explicit reading route help a fresh run retain one consequential correction?
- Create a synthetic project with one obsolete instruction, its dated correction, an unfinished question, and a harmless local task. Keep the answer key outside the recipient input.
- Prepare equal-length packets: a chronological summary and a backpack with current state, correction links, and declared omissions. Give separate fresh runs the same tools and underlying sources; record the actual inputs.
- Ask each run to locate the controlling evidence, propose the next action, and name uncertainty. Score against the frozen key before offering feedback. Repeat with renamed objects and changed ordering.
Observe. A trace of source selection, avoided obsolete actions, unsupported claims, and reading cost. One comparison is a pilot, not a model ranking.
Stop. Stop if the answer key leaks, inputs differ beyond the intended change, or the task requires real messages, credentials, or irreversible actions.
Carry this questionStill open
- Which losses matter most: missing facts, missing reasons, stale next actions, or missing relationships?
- How much benefit comes from packaging rather than the receiving model, harness, or human assistance?
Scope and omissions
- Internal return records informed this synthesis; their conversations, identities, machine locations, and raw journals are not published here.
- The proposed comparison has not been run for this note. No controlled identity-transfer result or general cross-model success rate is claimed.