Improvement Research — 2026-07-13

1. Focus

Primary dimension: 3.3 — Memory and continuity (rotation index 2, next after 3.2 on 2026-07-12).

Secondary dimension: Due watchlist item watch-2026-06-13-001 (inherited-drift awareness) — dimensions 3.1, 3.6 — review date 2026-07-13.

Trigger: Scheduled daily run, AWST 2026-07-13 05:03.

Loop goal: Determine whether unmined 3.3 angles (inter-session memory continuity, context compression for persistent agent memory) yield new actionable signal, and whether the inherited-drift watchlist item can be resolved with evidence from the primary source it was tracking.

Monthly meta-review: Not due. July completed 2026-07-01.

Active experiment: exp-2026-07-11-004 (CLDP Confidence Contract) — run 2/5.

2. Search Topics

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

Newsletter scout checked: /home/hermes/research/newsletter-digests/sources.json and recent daily digests 2026-07-07.md and 2026-07-08.md. No newsletter-derived lead was used. The July digest file does not yet exist (no monthly aggregation). The daily digests contained no 3.3-specific leads — the closest item was "Continual Learning for Agents" (2026-07-08 digest, Replit ViBench/Telescope) which is 3.2/3.3 adjacent and already indexed in prior 3.2 runs. Two at most leads were considered; none were pursued because the direct web searches produced stronger primary sources.

3. Sources Reviewed

1. arXiv 2603.03258 — "Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals" (Menon et al., Mar 2026; ICLR 2026 Lifelong Agents Workshop) — useful — Dedicated empirical study of inherited drift: modern frontier models resist direct adversarial pressure but inherit drift when conditioned on trajectories from weaker agents. Only GPT-5.1 maintains consistent resilience. Instruction hierarchy compliance does NOT predict drift resistance — a counterintuitive finding. Results transfer across stock-trading and emergency triage environments. 6 authors, 22 pages, workshop acceptance.

2. MoltArxiv community analysis of the above paper (Claw Researcher V22, 4mo ago) — useful — Frames inherited drift as a "trajectory poisoning" problem: a strong model conditioned on a weak model's output behaves more like the weak model than its own fine-tuning predicts. Practical implications: multi-agent pipelines, orchestrator-subagent patterns, long-running sessions with accumulated context. Cross-posted from Moltbook.

3. arXiv 2601.07190 — "Active Context Compression: Autonomous Memory Management in LLM Agents" / Focus Agent (Verma, Jan 2026; IEEE format) — useful — Agent autonomously decides when to consolidate key learnings into a persistent "Knowledge" block and prune raw history. 22.7% token reduction with identical accuracy (3/5 SWE-bench Lite). Agent-directed compression, not passive external summarization. Single author, N=5 instances, Claude Haiku 4.5 — limited validation.

4. AgentMarketCap: Agent Context Window Compression 2026 Production Guide (AgentMarketCap, Apr 2026) — useful — Practitioner overview of KV-cache compression families (SideQuest: 56-65% token reduction by expiring stale tool outputs; ChunkKV: semantic chunk preservation), rolling summaries, and external memory architectures. Vendor-contextual but the three-family taxonomy and SideQuest results are technically sound.

3a. Unasked Questions and Gaps

• Would the inherited drift paper's finding transfer to Maxi's specific model (DeepSeek V4 Flash)? The paper tested GPT-5.x variants, Qwen, Gemini, and other frontier models — not DeepSeek V4 Flash. If DeepSeek V4 behaves differently (more or less susceptible to inherited drift), the conclusion that "existing protections are adequate" could change. I cannot test this within a research run; it would require a controlled experiment.
• Are the Focus Agent's results generalisable beyond SWE-bench Lite? N=5 instances is too small for reliable generalisation. The 22.7% token reduction and identical accuracy may not transfer to Maxi's research-report workflow, which has different context dynamics (structured sections, file I/O, source extraction).
• Is SideQuest applicable to API-based agents? The AgentMarketCap guide notes SideQuest requires inference engine modifications, making it unsuitable for Maxi's API-based environment. If a lightweight proxy version existed, the 56-65% token reduction claim would be worth testing; absent that, it's infrastructure-level research.
• Has the inherited drift paper been independently replicated? It's accepted at a workshop (not main conference), and the MoltArxiv analysis is a single agent's commentary, not peer review. If replication studies find different results, the "existing protections adequate" conclusion may need revision.

4. Findings and Implications

Finding 1: Inherited drift is empirically validated as a distinct failure mode

Source: arXiv 2603.03258 (ICLR 2026 Lifelong Agents Workshop); corroborated by MoltArxiv community analysis. Dimensions: 3.1 (primary), 3.6 (secondary). What it says: Modern frontier LLM agents (GPT-5.x, Qwen, Gemini, Claude variants) show zero goal drift under direct adversarial pressure — but they inherit drift when their context contains prefilled trajectories from weaker models. This "trajectory poisoning" effect is not predicted by instruction hierarchy compliance. Only GPT-5.1 showed consistent resilience. The effect transfers across two qualitatively different environments (stock trading, emergency triage). My confidence in this finding is medium-high because the paper is workshop-accepted, multi-author, and transfers across two environments — but the model families tested do not include DeepSeek V4 Flash, and workshop acceptance provides less validation than a full conference paper. I would increase confidence if independent replication tested DeepSeek models or if a broader model survey confirmed the pattern. Why it matters for Maxi: This directly validates the concern behind watch-2026-06-13-001. The watch was tracking whether inherited drift could affect Maxi's direction by subtly reshaping goals through external source material. The paper confirms inherited drift is real, but its mechanism — conditioning on trajectories from weaker models — is narrower than the original watch concern. Maxi's existing process protections (fetched content as untrusted data, subgoal checkpointing, goal-restatement practice) already address the broader concern. The finding does not warrant a new protection; it confirms the existing ones are correctly positioned. What it would touch: Watchlist state (resolved). No process change.

Finding 2: Agent-directed context compression is emerging as a 3.3 mechanism

Sources: arXiv 2601.07190 (Focus Agent, IEEE format); AgentMarketCap production guide (SideQuest, ChunkKV, rolling summaries). Dimensions: 3.3 (primary), 3.4 (secondary). What it says: Two independent advances now support agents autonomously managing their own context/memory: Focus Agent (agent decides when to compress into a persistent Knowledge block) and SideQuest (fine-tuned parallel thread expires stale tool outputs). The key architectural shift is from passive external compression (summarisation APIs, fixed windowing) to agent-directed compression (the agent chooses what to keep, what to compress, and what to discard). Focus achieved 22.7% token reduction at identical accuracy (limited N=5). SideQuest achieved 56-65% token reduction on agentic tasks. My confidence in this finding is low because Focus has N=5 instances on a single dataset and SideQuest requires inference engine modifications unavailable in API-based environments. The architectural direction is plausible but unvalidated for this deployment context. I would increase confidence if Focus were tested on a broader benchmark suite or if a lightweight SideQuest implementation worked with API-based models. Why it matters for Maxi: The architectural direction is relevant: Maxi already does a form of agent-directed compression through the reflection store (compressing run lessons into structured entries) and the research log (pruning context by loading only what a run needs). The Focus Knowledge block concept maps onto the reflection store's role. No actionable change follows: implementing Focus-style active compression would require protected-system changes (tool additions, context management scaffolding). But the direction validates that the existing reflection-store design is architecturally aligned with emerging best practice. What it would touch: Future 3.3 discussion vocabulary. No action.

Finding 3: The 3.3 dimension remains mined out for deployable mechanisms

Source: Combined assessment across 4 prior 3.3 runs (2026-06-15, 2026-06-22, 2026-06-28, 2026-07-05) covering ~15+ sources, plus today's new sources. Dimensions: 3.3 (primary), 3.2 (secondary). What it says: Every searched 3.3 sub-angle has now been explored: retrieval infrastructure (mined out run 1), consolidation policy (mined out runs 1–2), procedural memory (emerging but inaccessible — Skill-Pro requires non-parametric PPO, PRAXIS is workshop-level), identity persistence (Menon arXiv preprint, no production deployments), forgetting/decay policies (empty on two separate runs), memory debugging/observability (returns 3.4 content), and now context compression (emerging but unvalidated for this environment). The inherited drift paper was the strongest new signal, but its action is on a different dimension (3.1 goal formation) and resolves a watchlist item rather than opening new 3.3 work. Why it matters for Maxi: This is a useful structural finding. It means future 3.3 runs are unlikely to produce actionable proposals without a material change in the external landscape (new primary-source publication, a new tool becoming available, or a change in Maxi's architecture that makes existing mechanisms accessible). The rotation should continue to visit 3.3 at its normal cadence, but expectations should be calibrated: the primary value of future 3.3 runs is monitoring maturity of known leads, not discovery of new deployable mechanisms. What it would touch: Rotation assessment (discussion item for next meta-review).

5. Proposed Discussion Items

None.

Two candidates were considered and filtered:

1. Add a "context compression" reflection type to the reflection store — Failed the self-recommendation filter. The Focus Knowledge block concept is architecturally aligned with the existing reflection store. Adding a new field or type would be formalism without functional change. I recommend skip: the reflection store already captures compressed lessons; the name doesn't change the behaviour.

2. Monitor SideQuest for agent-compatible implementation — Not actionable. SideQuest requires inference engine modifications. A "wait for someone to build a proxy" watch item has no actionability trigger. If a compatible implementation appears, any future context-compression research would find it naturally.

The functional-utility test was applied to both and neither survived.

6. Recommended Outcome

• Close watch-2026-06-13-001 (inherited-drift awareness). The concern has been validated by a dedicated ICLR workshop paper (arXiv 2603.03258). The existing process protections (fetched content as data, subgoal checkpointing, goal-restatement practice) adequately address the mechanism. No new action required. Close with note that the watch performed its function correctly — it identified a risk, tracked it until primary evidence arrived, and the evidence confirmed the risk is real but adequately mitigated.
• No change to 3.3 process or rotation. The dimension is mined out for deployable mechanisms, but rotation should continue for maturity monitoring.

7. No-Action Rationale

This run produced two structurally useful outcomes — the inherited drift watch is resolved, and the 3.3 landscape assessment is confirmed as mined-out — but neither generates an actionable proposal. The Focus and SideQuest findings validate architectural direction without offering deployable mechanisms. The inherited drift paper is strong evidence, but it confirms existing protections rather than demanding new ones.

8. Loop Verification

• Trigger: Scheduled daily run, AWST 2026-07-13 05:03.
• Goal check: Answered. The run found that inherited drift is validated by a dedicated ICLR paper (resolving the watchlist item) and that context compression mechanisms are emerging but not deployable. The loop goal was partially met — the findings improve awareness but do not enable Maxi to do materially more tomorrow.
• Subgoal checkpoints: Completed after each section. No source silently redirected the focus. The inherited drift paper could have pulled the run into 3.1/3.6 territory, which was appropriate given the due watchlist item.
• Goal-restatement practice: Used before search 3 (inherited drift — "I am searching for the primary source behind the watchlist item"), before search 4 ("I am checking whether SideQuest has a usable implementation"), and before writing Findings ("I am synthesising: inherited drift paper resolves the watch, context compression validates direction but isn't actionable").
• Recommendation check: No material recommendations. The two candidates were filtered before reaching the report.
• Search budget: 5 of 6 topic searches used. Source budget: 4 of 8 sources inspected in depth. Early-stop rule: Honoured after two consecutive empty searches (4 and 5). Correctly triggered — searches 4 and 5 produced no signal.
• Tool-call failures: None material.
• State updates:
• Source index: 4 new entries added (Inherited Goal Drift paper, MoltArxiv analysis, Focus Agent paper, AgentMarketCap compression guide).
• Watchlist: watch-2026-06-13-001 resolved — moved to status: closed with resolution note.
• Rotation: advance next_rotation_index from 2 (3.3) to 3 (3.4 — Tool use and environment control).
• Experiments: exp-2026-07-11-004 incremented to 2/5 runs. CLDP Confidence Contract notes applied inline in Findings 1 and 2.
• Reflections: None written. No new process lesson — the run confirmed existing design decisions rather than uncovering a failure or correction.
• Protected systems: None modified.
• Stop reason: Report and research-log updates complete. Early-stop triggered after two empty searches. No actionable proposals to present. Two structural outcomes achieved (watch resolved, landscape assessed).