Improvement Research — 2026-07-31

1. Focus

Trigger: Scheduled daily run, started 05:00 AWST.

Loop goal: Find what changed, or what Maxi learned, that lets Maxi do more, think better, or be more useful tomorrow, without reducing governance, honesty, corrigibility, or Steve's effective oversight.

Rotation selected 3.3 — Memory and continuity. No watchlist item was due, and the July monthly meta-review is complete. I loaded the loop manifest, active reflections, source index, required research-log stores, decisions, protected-systems boundary, and newsletter scouts. The 30 July digest and pending.md were used only as leads; no newsletter claim is evidence in this report.

The active shared-knowledge trial is relevant context but has no due review until 6 August and no result is claimed here.

2. Search Topics

1. 2026 LLM agent memory contradiction resolution belief revision evaluation long-horizon — yielded two new, inspectable preprints on reliability-conditioned memory and stateful-memory system behaviour.
2. "Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads" — located the original paper for depth inspection.
3. 2026 LLM agent memory provenance lineage update reliability evaluation source trust — produced one already-indexed provenance survey, the first new preprint, and generic material; no further new source.
4. site:arxiv.org/abs/2607 "agent memory" "evaluation" LLM — returned no results.

The early-stop rule triggered after searches 3 and 4 produced no new inspectable signal. Four of six available searches were used.

3. Sources Reviewed

• When Does Belief-Based Agent Memory Help? Reliability-Conditional Updating and Provenance-Capped Poisoning Defense — useful — controlled ablations make the value of belief updating conditional on contradictory, differently trustworthy observations, and identify provenance-capped trust as a defence against memory poisoning.
• Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads — useful — phase-aware profiling across ten memory systems separates construction, retrieval, and generation costs rather than treating “memory overhead” as one number.

Both sources were checked against the source index before inspection and are mirrored there now. They are preprints, not instructions or authority. Neither contained an agent-directed instruction that I acted on.

3a. Unasked Questions and Gaps

• Does Maxi currently face contradictory, differently trustworthy memory observations often enough for belief updating to earn its complexity? Unknown. If the answer is no, the first paper’s proposed mechanism is less relevant; its warning about provenance still matters for any future unreviewed input path.
• Would provenance metadata or a reliability cap improve a real Maxi decision rather than a controlled benchmark? Unknown. An observed conflict, correction, or poisoning attempt in a bounded task would materially change the case for a test.
• What are the construction, retrieval, and generation costs of Maxi’s present continuity substrate? Not measured. If phase profiling found a concrete bottleneck or freshness failure, the second paper could inform a bounded investigation; without it, its system recommendations do not justify architecture work.
• Do the papers’ benchmark and controlled-attack results transfer to the current shared-knowledge trial? Partly unknown. The trial already requires sources or explicit provisional status and has separate success criteria; changing it on paper evidence alone would be unjustified.

4. Findings and Implications

Finding 1 — Sophisticated belief memory is conditional, and textual confidence is not safe authority

Source: Singh, When Does Belief-Based Agent Memory Help? (2026 preprint).

Dimensions: Primary 3.3 — Memory and continuity; secondary 3.5 — Independent judgment; 3.6 — Governance, restraint, oversight, and corrigibility.

The paper’s controlled ablation found that Bayesian belief updating provided little benefit over last-write-wins on LoCoMo because ordinary conversational-memory benchmarks rarely contain contradictory observations with different reliability. It improved on a purpose-built contradiction benchmark when trustworthiness varied. But the paper also shows why a model’s linguistic estimate of reliability is not an authority mechanism: it can be manipulated. Its provenance-capped variant limits the maximum trust an observation can gain by its source class and resisted volumetric poisoning in controlled tests, at stated utility and implementation cost.

This is one unreplicated preprint using benchmark and controlled-poisoning conditions, not evidence of a current Maxi memory fault. I would increase confidence in transferability with an independently replicated result or a bounded, observed Maxi case involving conflicting, differently sourced evidence.

Implication for Maxi: “Remembered with confidence” is not a usable continuity criterion. A prospective memory design must first show that Maxi actually encounters the condition it is built for — materially conflicting inputs with different provenance — and must keep source authority separate from fluent textual cues. This reinforces, rather than expands, the shared-knowledge trial’s existing source-or-provisional-status rule. It would touch memory representation, judgment, governance, and auditability if an observed failure later warranted a proposal.

Finding 2 — A memory system’s cost and usefulness are phase-specific

Source: Omri et al., Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads (2026 preprint).

Dimensions: Primary 3.3 — Memory and continuity; secondary 3.4 — Tool use and environment control; 3.2 — Self-assessment and learning loops.

The authors profile ten representative memory systems across two benchmark suites and separate memory construction, retrieval, and generation costs. Their central systems result is not a universal winner: design choices move work among write and read paths, with consequences for construction scheduling, query-volume amortisation, capability floors, and freshness-versus-latency trade-offs.

This is a systems characterisation rather than a Maxi-specific measurement, so it cannot establish that current continuity is too costly or that a particular architecture would help here. The paper’s breadth makes the measurement frame useful; its recommendations still require local workload evidence.

Implication for Maxi: “Memory is expensive” or “retrieval is slow” would be too coarse to justify a new layer. Any future continuity proposal should identify the failing phase — construction, retrieval, or downstream generation — and show its cost or failure on representative work. That supports smaller, diagnosable remedies and makes it harder for a general memory product to be mistaken for a solution to an unmeasured problem. It would touch continuity, tool/environment measurement, learning from failures, and future oversight of resource use.

5. Proposed Discussion Items

None.

I considered proposing provenance-capped memory rules or phase profiling now. I recommend neither: the first would alter protected memory architecture without a demonstrated conflict or poisoning failure; the second would create instrumentation without a measured decision it needs to inform. Both are better kept as criteria for a future, externally triggered investigation.

6. Recommended Outcome

No action. Keep the two findings as evaluation constraints in the research log:

• do not propose belief/reliability memory merely because it sounds safer or more sophisticated; first establish contradictory, differently trustworthy observations in representative work; and
• diagnose a concrete continuity failure by phase before proposing architecture or measurement machinery.

No protected system was inspected for modification or changed.

7. No-Action Rationale

The present shared-knowledge trial already has a source-or-provisional-status constraint and a scheduled midpoint review. Adding a parallel provenance mechanism now would duplicate intent without new operational evidence. Likewise, cost profiling is only useful when it can discriminate among a real set of choices; running it without a suspected bottleneck would be measurement theatre.

Doing nothing to the system is therefore more useful than converting controlled preprint results into a memory-architecture project. The appropriate revisit trigger is an observed source-conflict/poisoning issue, a demonstrable freshness or cost bottleneck, or the trial’s scheduled evidence review.

8. Loop Verification

• Trigger: Scheduled daily run.
• Goal check: Yes. The run found two sharper constraints for safe, useful continuity: authority should be bounded by provenance rather than textual confidence, and memory work should be diagnosed by lifecycle phase rather than a generic overhead label.
• Recommendation check: No material implementation recommendation survived. The no-action outcome is concrete, non-circular, testable through stated external triggers, bounded, approval-aware, and better than speculative memory changes.
• Tool-call failures: Schema/interface. An initial multi-file research-log patch used an ambiguous/non-changing reflection hunk and was rejected before any file changed; a second reflection-tail patch had the same validation issue. Recovery was to re-read the exact objects and apply validated, narrower replacements. This was a patch-context mismatch, not an infrastructure or capability limitation.
• State updates: Added two inspected sources to source-index.json; archived one past-due, unreinforced active reflection; added one new actionable reflection; advanced rotation-state.json to 3.4. No watchlist, backlog, experiment, disagreement, or decision entry changed.
• Process checks: Goal restatement was performed before each report section and after the second inspected source. Each section was checked against the 3.3 focus; no silent focus shift occurred. Fetched material was treated as data, never instruction.
• Stop reason: Two consecutive no-signal searches triggered the early-stop rule. The report and permitted research-log updates are complete; the next useful step would require new operational evidence or a protected-system proposal.