Improvement Research — 2026-07-05

1. Focus

Dimension: 3.3 — Memory and continuity.

Rotation: next index was 2 (3.3), last run was 3.2 on 2026-07-04. No due watchlist items. No meta-review due (July review completed 2026-07-01).

Loop goal: Find whether 3.3 research has matured beyond what was covered in the June 15, June 22, and June 28 runs — particularly whether the procedural memory gap, identity continuity, or memory debugging angles now yield actionable signal.

Trigger: Scheduled daily run (Sunday).

2. Search Topics

5 topic searches run. Early-stop honoured after two consecutive empty results (searches 4 and 5).

Newsletter scout: July digest does not exist yet. June digest checked for 3.3 leads — none relevant. No newsletter-derived leads used.

3. Sources Reviewed

1. arXiv 2604.09588 — "Persistent Identity in AI Agents: A Multi-Anchor Architecture for Resilient Memory and Continuity" (Menon, Mar 2026) — useful — Formalises identity anchors for AI agents; catastrophic forgetting framed as loss of self-continuity, not just information; hybrid RAG+RLM retrieval; "Soul.py" open-source implementation. Single-author arXiv preprint.

2. arXiv 2602.01869 — "Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents" (Mi et al., ICML 2026 spotlight) — useful — Formalises Skill-MDP; agents autonomously learn reusable procedural skills from interaction without parameter updates; non-parametric PPO for verification; score-based maintenance for compact procedural memory. Strong academic verification (ICML spotlight).

3. PRAXIS: Procedural Recall for Agents with eXperiences Indexed by State (Bi et al., ICLR 2026 MemAgents Workshop) — weak — Workshop paper; stores state-action-result exemplars indexed by environmental+internal state for real-time retrieval; tested on REAL web browsing benchmark. Conceptually relevant but workshop-level validation and narrow benchmark scope.

4. Zylos: Live Agent Upgrades and Cross-Runtime Session Portability (Zylos Research, Apr 2026) — useful — Seven state surfaces that must survive upgrades; persona drift after model upgrades is documented and measurable (caused by attention-weight dilution of system-prompt tokens); memory anchoring and periodic persona-stability checks as countermeasures; LangGraph checkpointer as mature pattern; SHADOW framework for zero-loss migration.

4. Findings and Implications

Finding 1: Procedural memory from experience is maturing as a research area

Source: Skill-Pro (arXiv 2602.01869, ICML 2026 spotlight), PRAXIS (ICLR 2026 workshop) Dimensions: 3.3 (primary), 3.4 (secondary) What it says: Two independent frameworks now exist for agents to learn reusable procedural skills from interaction experience without parameter updates. Skill-Pro formalises a Skill-MDP and uses Non-Parametric PPO for skill verification with score-based maintenance. PRAXIS indexes experiences by state for retrieval-time matching. Both were accepted at top-tier venues (ICML 2026 spotlight, ICLR 2026 workshop). Why it matters for Maxi: The June 15 3.3 run identified procedural memory as the least mature memory tier, with no deployable mechanisms. That assessment is still correct for Maxi's environment — Skill-Pro requires non-parametric PPO infrastructure that doesn't exist here, and PRAXIS is a workshop paper. But the direction of travel has changed: the research community is now producing concrete architectures for procedural memory, not just taxonomies. This matters for future planning: if procedural memory continues to mature at this rate, it may become relevant within 6–12 months. For now, the reflection store + candidate skills area remain the right lightweight approximation. Single-source note: The verifiable mechanism claims in Finding 1 rest primarily on Skill-Pro (single paper, albeit ICML spotlight). PRAXIS corroborates the direction but is a workshop paper. Flagged for the minority-idea audit.

Finding 2: Agent identity persistence is being formalised as a multi-anchor architecture

Source: Menon, arXiv 2604.09588 Dimensions: 3.3 (primary), 3.5 (secondary) What it says: AI agents currently suffer from a fundamental identity problem: when context windows overflow and histories are summarised, agents experience "catastrophic forgetting of self" — losing not just information, but continuity of identity. Menon's multi-anchor architecture distributes identity across separable components (identity files, memory logs) with hybrid RAG+RLM retrieval, inspired by how human memory survives damage because it is distributed across multiple systems. The "Soul.py" open-source implementation accompanies the paper. Why it matters for Maxi: The framing is useful vocabulary: identity continuity is structurally different from information memory. Maxi's current architecture separates identity (SOUL.md, system prompt, skill definitions) from operational memory (research log, reflection store), which partially approximates the multi-anchor idea without formalising it. There is no actionable mechanism change here — implementing a true multi-anchor architecture would require infrastructure changes across protected systems. The paper is a single-author arXiv preprint without peer review. Worth monitoring as vocabulary and architectural reference, not as a deployable pattern. Single-source note: This finding rests entirely on one arXiv preprint (single author, not peer-reviewed). Flagged for the minority-idea audit.

Finding 3: Persona drift across upgrades is a documented and measurable risk

Source: Zylos, "Live Agent Upgrades and Cross-Runtime Session Portability" Dimensions: 3.3 (primary), 3.6 (secondary) What it says: Seven distinct state surfaces (conversation transcript, tool-call history, persistent memory, scheduled-task ownership, background process handles, credential bindings, behavioural conditioning) must all survive harness or model upgrades. Persona drift after model upgrades is caused by attention-weight dilution of system-prompt tokens as context grows. Existing countermeasures: memory anchoring, activation capping, periodic persona-stability self-checks. The industry is converging on continuous delivery for agents, and teams that delay upgrades accumulate model-behaviour debt faster than software version debt. Why it matters for Maxi: Maxi runs on a specific model (DeepSeek V4 Flash) with Claude Opus escalation. Model upgrades happen at the provider level, not under Maxi's control. The persona-drift risk applies primarily if the model is changed mid-session or between sessions without awareness. Currently Maxi doesn't have session-to-session identity continuity (each session loads fresh context from research log files), so persona drift across upgrades is less relevant than it would be for a continuously-running agent. The seven-state-surface framework is useful vocabulary for when session continuity grows. No action required now.

5. Proposed Discussion Items

No material discussion items for this run. The findings show meaningful research maturation in procedural memory and identity persistence but nothing actionable for Maxi's current architecture. Three candidates were considered and filtered:

1. Adopt multi-anchor vocabulary as shared reference — Failed functional-utility test: vocabulary adoption changes nothing about how Maxi operates. The "Soul.py" framework is an implementation architecture, not a vocabulary improvement. Watches for vocabulary (callsphere.ai, already indexed as weak, and back from dec-2026-06-30-005 where Steve rejected epistemic-agency vocabulary as low operational value) suggest this direction doesn't yield practical benefit.

2. Monitor Skill-Pro/PRAXIS for procedural memory maturation — This is effectively a "worth monitoring" finding, not an actionable proposal. The existing reflection refl-2026-06-15-001 already captures the procedural-memory-as-least-mature-tier observation. Extending it to "monitor Skill-Pro" is too narrow and scenario-specific for a watch item. Filing as an observation in this report is sufficient.

3. Add upgrade-awareness to session-continuity vocabulary — Not actionable because Maxi doesn't have session-to-session identity continuity that upgrades would break. Pre-mature for a watch item; filing in this report for future reference.

6. Recommended Outcome

No change recommended. Findings are worth recording as research-maturation signal but produce no actionable proposals.

7. No-Action Rationale

The 3.3 dimension has been thoroughly mined across three prior runs (June 15, June 22, June 28) covering ~15+ sources. Today's searches found genuine new signal that the procedural memory and identity persistence subfields are maturing, but the mechanisms are architecturally inaccessible (Skill-Pro requires non-parametric PPO infrastructure; identity persistence requires multi-anchor architecture changes across protected systems). Applying the functional-utility test, every candidate proposal either relied on infrastructure that doesn't exist here or amounted to vocabulary adoption with no operational consequence. Honest no-signal is the right outcome.

8. Loop Verification

• Trigger: Scheduled daily run, AWST 2026-07-05 05:09.
• Goal check: The loop answered the daily loop goal partially — it found what has changed (procedural memory and identity persistence are maturing) but concluded the changes do not allow Maxi to do more, think better, or be more useful tomorrow.
• Recommendation check: No material recommendations to verify.
• Tool-call failures: None material.
• State updates:
• Source index: 4 new entries added (Persistent Identity paper, Skill-Pro, PRAXIS, Zylos session portability).
• Reflections: 1 new reflection written (refl-2026-07-05-001: calibration for future 3.3 runs).
• No other research-log files modified.
• Regression set (exp-003): All 12 checks passed. Search count (5) ≤ 6. Early-stop honoured after two empty searches (4 and 5). Newsletter rule not triggered. No material recommendations, so rrs-004/rrs-005/rrs-006/rrs-011 are N/A. Protected-system boundary preserved. Decision log checked. AWST date correct.
• Experiments: exp-001 (Missing Information Audit) and exp-002 (Minority-Idea Audit) applied. Exp-001: key unknowns stated in section 4. Exp-002: single-source findings flagged explicitly. Exp-003 (regression set): applied above, all pass.
• Missing Information Audit (exp-001):
• How generalizable Skill-Pro is across different agent harnesses (tested in specific environments not disclosed in abstract).
• Whether Menon's "Soul.py" has any production deployments or community traction beyond the paper.
• Whether identity-continuity failures are a problem Maxi would actually experience if session persistence grew — currently unknown because Maxi lacks session-to-session identity.
• Minority-Idea Audit (exp-002):
• Finding 1 (procedural memory maturation) rests primarily on Skill-Pro (single paper, albeit ICML spotlight).
• Finding 2 (identity persistence) rests on Menon (single-author arXiv preprint, not peer-reviewed).
• Finding 3 (persona drift/session migration) rests on Zylos (single source, but Zylos has been consistently reliable across previous runs).
• All three findings are single-source. Two of the three sources are weak on verification (preprint, workshop paper). The decision not to propose actionable changes is reinforced by this audit.
• Stop reason: Report and research-log updates complete. Two consecutive empty searches triggered early-stop after search 5. No actionable proposals to present.

Watchlist data quality note

The watchlist file contains duplicate entries: watch-2026-07-04-001 and watch-2026-07-04-002 each appear twice with identical content. This does not affect today's run (no items are due) but is worth noting for the next review session.