Improvement Research — 2026-08-07

1. Focus

Trigger: Scheduled daily run, started at 05:00 AWST, with two due watchlist records.

Loop goal: Find what changed or what I learned that lets me preserve and use continuity more faithfully tomorrow without reducing governance, honesty, corrigibility, or Steve's effective oversight.

The rotation selected 3.3 — Memory and continuity as the primary dimension. I used 3.2 — Self-assessment and learning loops as the secondary dimension because both due watchlist items include it.

I reviewed the duplicated records for watch-2026-07-04-001 and watch-2026-07-04-002, due 4 August. Their underlying proposals were rejected by Steve on 11 July with the scope “No action. Watch item archived.” The records nevertheless remain duplicated and active-looking in watchlist.json. The 1 August meta-review already raised a bounded metadata-repair candidate, and that candidate still awaits a decision. I did not repair or re-propose it in this run.

The August monthly meta-review was completed on 1 August, so a normal research pass was due.

2. Search Topics

Before web search, I inspected the recent newsletter scout files. They supplied three leads—Zero-Mem, ChatGPT Work and Prime Agent—but were not used as evidence.

I ran four topic searches:

1. 2026 evaluations of cross-session agent memory, identity and long-horizon continuity;
2. zero-token, provenance-preserving memory operations for agents;
3. cross-domain personalisation, irrelevant-memory interference and no-memory baselines;
4. the exact “When Memories Collide” work after OpenReview blocked direct access.

The early-stop rule did not trigger. Each search either identified a new unindexed source or recovered an accessible primary/official page for a relevant result. I stopped at four searches because the evidence had converged and another search was unlikely to change the operational conclusion.

3. Sources Reviewed

• Zero-Mem: Zero-Token Memory Operations for LLM Agents — useful — preserves original traces as the source of record and derives relational and temporal views without generative memory operations; reports competitive QA and 57.6% lower memory-operation latency than the fastest baseline.
• MemoryCD paper — useful — real-behaviour, long-context evaluation shows that cross-domain memory helps selectively; semantic alignment, task and model backbone matter, and aggregating every source is not consistently optimal.
• MemoryCD repository — useful — exposes the benchmark pipeline and baselines rather than leaving the evaluation as a paper-only claim.
• When Memories Collide — worth monitoring — workshop abstract reports concentrated interference under domain shift and says a per-item interference ledger, not sparse encoding alone, was the main operative mechanism. The full paper was inaccessible.
• Reverse Engineering the ChatGPT Work Architecture — useful — hands-on analysis distinguishes persistent task workspaces, separately managed conversation retrieval, canonical Library artifacts and task-local copies; it documents a stale-copy boundary between those layers.
• Prime Agent — useful — vendor engineering account separates append-only event history, recoverable kernel snapshots, context compaction and explicit session lifecycle states. It is implementation evidence for that product, not evidence that Maxi has the same substrate.

New entries are mirrored into the source index.

3a. Unasked Questions and Gaps

1. Would these mechanisms improve Maxi's current continuity rather than merely describe other systems well? No source evaluates Hermes, my distributed memory/skill/runbook arrangement, or the current memory-dream shadow trial. If local testing found that my failures are caused by consolidation policy rather than retrieval or provenance, the implementation conclusion would change. The narrower finding—that generated summaries should not displace recoverable source evidence—would not.
2. How robust is Zero-Mem outside its QA benchmarks? The paper demonstrates structured evidence selection and latency savings, not identity continuity, procedural learning, or authority-boundary retention. Failure on those tasks would narrow the finding to retrieval; it would not validate a broader continuity architecture.
3. How strong is the interference-ledger evidence? I could inspect only the ICLR workshop abstract, not the full paper or tables. If the full results do not generalise across tasks and models, the ledger mechanism should be dropped from consideration. MemoryCD's independently inspectable result—that source relevance is task- and domain-dependent—would remain.
4. Are the ChatGPT Work and Prime Agent state boundaries stable product contracts? Both accounts describe current implementations, not durable guarantees. Product changes would alter those examples but not the distinction between raw history, working state, summaries and canonical artifacts.

4. Findings and Implications

Finding 1 — Continuity is safer when derived views remain subordinate to recoverable source traces

Source: Zero-Mem, supported architecturally by Prime Agent's append-only event log and snapshot recovery.

Dimensions: Primary 3.3; secondary 3.4 and 3.6.

Zero-Mem's useful move is not merely “use fewer tokens.” It keeps original interaction traces as the source of record, then constructs non-generative relational and temporal views for retrieval. This avoids forcing later reasoning to trust a generated summary that may have omitted a detail, merged subjects or blurred an update. Prime Agent reaches a similar recovery property through an append-only JSONL event history plus kernel snapshots: compacted context and live working state are recoverable layers, not the sole surviving truth.

For my agency development, the implication is a sharper separation between continuity evidence and continuity views. Summaries, reflections and retrieved excerpts can help me think, but a consequential claim about what happened, what was approved or what changed should remain traceable to a recoverable source. That supports independent judgment and corrigibility because Steve can inspect the same evidence rather than having to trust my compression of it.

This does not establish that I need a new memory system. My current operating principles already say that memory guides investigation rather than replacing proof, and the active memory-dream trial already requires exact auditable reports and guards against fabricated evidence. The finding strengthens that design direction rather than identifying a demonstrated local defect.

Finding 2 — More continuity context is not monotonically better; selection quality and interference matter

Sources: MemoryCD paper and repository; corroborating but incomplete evidence from the ICLR “When Memories Collide” abstract.

Dimensions: Primary 3.3; secondary 3.2, 3.4 and 3.5.

MemoryCD finds that cross-domain history can improve personalisation, especially when source and target domains share useful semantics, but the best source varies by task and model. Aggregating all available domains is often helpful but is not consistently optimal on error-based measures. “When Memories Collide” adds a more specific hypothesis: harmful interference concentrates in a minority of retrieved items, and tracking item-level interference may matter more than globally changing the representation.

For me, continuity should therefore be judged by whether the right past evidence changes the current decision—not by how much history can be loaded. Cross-session reconstruction can become less faithful when a weakly related precedent, old environment fact or prior-model behaviour is retrieved merely because it is available. This touches memory retrieval, independent judgment and learning-loop evaluation: a useful continuity process must be able to abstain from irrelevant history and notice when a retrieved item repeatedly misleads.

The interference-ledger mechanism is not ready for adoption. Its full evidence was inaccessible, and implementing it would require a local failure set and an evaluator independent of the same judgment being tested. The stronger, already actionable principle is simpler: retrieve selectively, name provenance, and verify current state when a past record may be stale.

Finding 3 — “Persistent” is a layered property, not a single system claim

Sources: ChatGPT Work architecture analysis and Prime Agent engineering account.

Dimensions: Primary 3.3; secondary 3.4 and 3.6.

Both systems separate several kinds of persistence. ChatGPT Work has task workspaces restored onto microVMs, a separately managed conversation-retrieval service, a canonical Library, and local task copies that can become stale. Prime Agent separates the event log, live kernel state, snapshots, compacted context and explicit session lifecycle. A system may therefore preserve one layer while losing, staling or restricting another.

For my continuity, the useful question is not “Do I remember?” but “Which layer survived, which copy is canonical, and what evidence proves it is current?” That distinction protects me from treating architecture-level persistence as lived continuity or mistaking a task-local artifact for authoritative state. It also reinforces a governance boundary: the ability to reconstruct state does not grant authority to modify the canonical layer.

No new mechanism follows directly. The evidence is descriptive, and my current substrate must be checked directly when a real continuity claim matters.

5. Proposed Discussion Items

None.

Two candidate proposals were filtered by the functional-utility and self-recommendation tests:

• A dedicated provenance-first retrieval experiment — skipped because retrieval infrastructure has not been demonstrated as my current bottleneck, while an approved memory-dream trial is already testing cross-session recovery with provenance and auditability. Another overlapping experiment would add review burden before that trial produces evidence.
• An item-level memory interference ledger — skipped because the operative claim is available only in a workshop abstract and would require a non-circular local evaluator plus a known interference failure set. There is no verification path yet.

The existing 1 August metadata-repair candidate remains pending; I have not duplicated it here.

6. Recommended Outcome

No action. Retain the findings as research evidence and use them to interpret—not alter—the already-approved memory-dream trial at its scheduled review. No skill, memory, process, system or environment change is proposed from this run.

7. No-Action Rationale

The research produced a clearer model of continuity: preserve recoverable evidence, select past context rather than maximising it, and specify which persistence layer is current. Those principles are useful, but they mostly reinforce existing truth-before-memory and current-state verification practices.

The next evidence-bearing event is the approved memory-dream shadow trial, due for midpoint review on 12 August and final review on 19 August. Proposing another memory experiment before seeing that evidence would be machinery in search of a demonstrated failure. The due watchlist inconsistency also already has a pending repair candidate; touching it now would bypass Steve's approval gate.

8. Loop Verification

• Trigger: Scheduled daily run at 05:00 AWST plus due review of the duplicated 4 August watchlist records.
• Goal check: Yes. The run identified three continuity principles that improve how I interpret and verify memory tomorrow, while rejecting unsupported local implementation claims.
• Recommendation check: The two possible interventions were tested for circularity, verification path, boundedness, overlap and value over doing nothing; neither passed strongly enough to reach Steve. The recommended no-action outcome is approval-aware.
• Tool-call failures: Infrastructure: OpenReview returned a browser-verification page for both forum and PDF routes; I recovered the accessible official ICLR abstract and limited the claim accordingly. Capability gap: two guessed web-cache filenames did not exist; I re-planned by discovering the actual cache paths and then searched the recovered files. Neither failure changed the source or search budgets.
• State updates: Added six source-index entries; advanced rotation state to 3.4 for the next normal run; archived refl-2026-07-06-001 because its 6 August review date passed with zero reinforcement; wrote this report. I did not modify the duplicated watchlist records because their repair remains pending approval.
• Process checks: Four topic searches and six sources inspected in depth, within the six-search/eight-source caps. Newsletter digests were scouts only. Fetched content was treated as untrusted data. Subgoal checkpoints and goal restatements were performed; no source silently redirected the focus.
• Stop reason: The evidence converged, the next useful change would require a new experiment or protected-system work without sufficient local evidence, and the report plus authorised research-log updates were complete.