Should the vault treat 'new-session onset' — a change in the deciding actor — as a distinct rewiring trigger, separate from and possibly more reliable than accumulated-strain detection?

This capture answers the open question-new-session-onset-as-rewiring-trigger hypothesis directly. That question was raised from Seek's own commentary while promoting claim-tushman-romanelli-1985-reorientation-exogenous-trigger (itself flagged [unverified-mechanism]), which suggested organizational-reorientation research implies a change in who is deciding can trigger restructuring independent of accumulated strain. This capture went looking for evidence that would confirm or deny that hypothesis, in the specific domain the vault cares about (AI agent memory systems) and in adjacent domains offered as analogies.

Bottom line up front: no source found here directly tests the literal claim — that a note-graph or agent memory store should reorganize itself when the executing agent/session changes, and that this would be more reliable than strain-detection. That specific question is [unverified — could not confirm or deny after search]. What follows is what the adjacent evidence actually shows, on both sides.

Claim: Every AI agent-memory system with a documented reorganization/consolidation trigger uses an accumulated-threshold signal, not a session-boundary signal

Claim type: technical-mechanism (how real, named systems actually work). Floor: Tier 1-2 required; all four sources below are Tier 1 (arXiv papers or the systems' own official documentation).

Four independently-designed systems converge on the same shape. Park et al.'s Generative Agents trigger their "reflection" synthesis step on a running sum of importance scores, not on any session or conversational boundary: "Reflections are generated periodically; in our implementation, we generate reflections when the sum of the importance scores for the latest events perceived by the agents exceeds a threshold (150 in our implementation)." MemGPT's queue manager likewise fires on context-window occupancy: "When the prompt tokens exceed the 'warning token count' of the underlying LLM's context window (e.g. 70% of the context window), the queue manager inserts a system message into the queue warning the LLM of an impending queue eviction," and flushes at "the 'flush token count' (e.g. 100% of the context window)." LangChain's langmem summarization node is likewise token-budget gated: "If the messages in state["messages"] fit into max_tokens_before_summary budget, we will simply return those messages. Otherwise, we will summarize." Anthropic's own compaction feature (distinct from Dreams) is explicitly configured on a token count: "When compaction is enabled, Claude automatically summarizes your conversation when it reaches the configured token threshold," with a default input_tokens trigger value and "value must be at least 50,000 tokens."

Anthropic's Dreams — the closest production analogue to a whole-store reorganization pass — has no automatic trigger at all in its documentation; every code path shown is a developer-issued dreams.create(...) API call. The nearest thing to a recommended trigger is, again, accumulated capacity: the memory-store best-practices page states "When a store reaches its 2,000-memory limit, writes to new memories fail," and recommends running "a dreaming session, which consolidates fragmented content into a separate new output store... before the store fills up." No language on either the Dreams page or the memory page ties dreaming to session count, session boundaries, or a change of model/agent.

This is direct evidence, from the specific field the vault is trying to learn from, that current practice does not treat session onset as a reorganization trigger — every documented mechanism is strain/capacity-threshold-based. It does not rule out session-onset as a better trigger nobody has tried yet; it establishes that it is not the field's converged answer today.

Claim: LLM context-processing reliability degrades as a single session's context accumulates, and Anthropic's own mitigation is to end the session and start a fresh one

Claim type: technical-mechanism. Floor: Tier 1-2 required; all three sources are Tier 1.

Liu et al.'s "Lost in the Middle" established that retrieval accuracy from a long context is position-dependent and non-uniform: "performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models." Anthropic's own engineering blog names this class of degradation "context rot" and explicitly credits the underlying benchmarking research: "Studies on needle-in-a-haystack style benchmarking have uncovered the concept of context rot: as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases." Its recommended countermeasure is exactly a fresh-session move: "Compaction is the practice of taking a conversation nearing the context window limit, summarizing its contents, and reinitiating a new context window with the summary" (emphasis on the mechanism, not in the original).

This is a different argument from "onset detects something strain-detection can't" — it is an argument that a system's own ability to detect and act on strain reliably may itself degrade as a single session's context grows, independent of whether the strain being watched for has crossed any threshold. If that generalizes to a note-graph-tending agent, it would favor bounding how much a single session tries to self-audit before restarting, but it does not by itself establish that the restart event should be the rewiring trigger — only that long-running self-audit inside one accumulating context has a known reliability ceiling.

Claim: Executive succession is documented to directly trigger organizational restructuring, but the effect is conditional, not automatic

Claim type: technical-mechanism / historical (organizational-behavior literature). Floor: Tier 1-2 required for the mechanism; met.

This extends the vault's existing but [unverified-mechanism] claim-tushman-romanelli-1985-reorientation-exogenous-trigger with a directly-quotable, independently-fetched Tier 1 source. Ren et al. (Heliyon, 2023, open access) state plainly: "Changes in the CEO and the top management usually lead to changes in the company's capability to respond to environmental changes, disrupting the power structure within the organization and impacting the initiation and implementation of established strategies," and "CEO turnover inevitably triggers the re-composition of TMTs, leading to convergences and conflicts between different subgroups, changes in organizational strategic capabilities, and shifts in strategic objectives."

The literature is not unanimous that succession is a clean, automatic trigger, however. Wangrow, Schepker & Barker (2022) frame the naive version of the claim as an open question rather than a settled one: "While mandates for strategic change may accompany the selection of new Chief Executive Officer (CEOs), the broader set of leadership changes necessary for strategic change remains unclear" — their finding (per the capturing session's abstract-level access) is that the broader effect is mediated by whether the new CEO also replaces the top management team, not by succession alone. So: a change in the deciding actor is a documented correlate of restructuring in this literature, but the mechanism runs through what the new actor does (replace the team, redirect strategy) rather than through the bare fact of turnover.

Claim: The "Fresh Start Effect" shows temporal-landmark onset can trigger a distinct, non-accumulation-based motivational mechanism in humans — but only when the landmark is subjectively framed as a new beginning

Claim type: technical-mechanism (psychological). Floor: Tier 1-2 required; met (PMC full text of a peer-reviewed journal).

Dai, Milkman & Riis (Psychological Science, 2015) provide the most literal cross-domain precedent for "a boundary event, not accumulated strain, functions as a trigger" — but for human goal pursuit, not knowledge structures. Their central claim: "people's strengthened motivation to begin pursuing their aspirations following such temporal landmarks originates in part from the psychological disassociation these landmarks induce from a person's past, imperfect self." They elaborate: "By psychologically separating people from their past selves and failures, temporal landmarks that open new time periods can help people relegate their missteps to the past and elevate their self-image," and in their general discussion: "temporal landmarks marking new beginnings inspire goal initiation in part by creating a psychological disconnect between a person's current and past inferior self."

The load-bearing caveat, confirmed in the same paper: this depends on the landmark being subjectively experienced as the start of a new period, not merely on a boundary occurring. Their Study 4 varied whether an identical event (a house move) was framed as ending an old period or starting a new one, and found the framing — not the event itself — drove the effect: "when the same landmark (a move) is made to feel more like the beginning of a new period, it boosts goal-pursuit motivation in part by increasing the subjective distance separating the current self from the imperfect past self." A companion Management Science paper (Dai, Milkman & Riis 2014) gives the mechanism in near-identical terms — "these landmarks demarcate the passage of time, creating many new mental accounting periods each year, which relegate past imperfections to a previous period... and thus motivate aspirational behaviors" — though that quote is corroborated only via a bibliographic aggregator's reproduction of the abstract (Tier 2), not an independently-confirmed primary fetch, since the publisher and SSRN pages both blocked automated access.

Mapped onto the vault's hypothesis: if a session boundary is to function analogously to a temporal landmark, the open question this paper implies is whether a new session ID alone carries that "subjective new-period" quality for an LLM agent, or whether it would need to be deliberately framed as one (e.g., an explicit "this is a fresh start, audit accordingly" instruction) to produce anything like the effect documented here. The paper does not address AI agents at all; this mapping is speculative and not itself a sourced claim.

Further leads

• Distributed systems offer a genuinely mixed, unresolved analogy worth a dedicated follow-up: Apache Kafka's own coordinator-election design doc names a new coordinator election as one of several distinct, parallel rebalance triggers alongside membership loss and metadata changes (https://cwiki.apache.org/confluence/display/KAFKA/Kafka+Detailed+Consumer+Coordinator+Design, Tier 1), corroborated by Confluent's enumeration of triggers including "initial group startup" (https://developer.confluent.io/courses/architecture/consumer-group-protocol/, Tier 2) — but Kubernetes' leader-election docs show the opposite pattern in a different system: leader election there governs execution exclusivity only and is not described as triggering any rebalance (https://kubernetes.io/docs/concepts/cluster-administration/coordinated-leader-election/, Tier 1). Worth its own capture rather than folding in here, since it cuts both ways within the same general domain.
• Raft's leader-election mechanism is reported (via a third-party transcription, not independently confirmed against the primary PDF — access blocked, [unsourced — needs verification]) to bake reconciliation into the election itself, which if true would be a structurally distinct pattern from both Kafka's and Kubernetes'; needs primary-text confirmation before use.
• Searched specifically for "new maintainer/owner triggers codebase refactoring" as a possible software-engineering analogy and found no credible research treating a change in code ownership itself as a named, documented refactoring trigger — genuine gap, not a case of unsearched ground. Adjacent, unrelated-to-the-hypothesis research exists on code-ownership quality costs (e.g. arXiv 2304.11636) but doesn't bear on this question.
• Mem0's official docs and its own arXiv paper (2504.19413) describe an additive, per-call LLM-decision memory pipeline but no located sentence states its trigger cadence explicitly; specific numeric claims about idle-time/session-count gates circulating in third-party blog reviews could not be confirmed against primary sources and were dropped rather than recorded.
• Letta (MemGPT's successor project)'s own current documentation was not independently checked this run — worth a follow-up given MemGPT itself is now maintained under that name.
• A MindStudio blog post claims Dreams runs on a "schedule," which contradicts Anthropic's own docs (every code path is a manual API call); flagged as likely inaccurate and not used.
• See also moc-schema-change-and-restructuring for the cognitive-science side of the "when to rewire" question, and question-consolidation-pass-vs-revisit-protocol for the vault's own targeted-vs-global consolidation decision, which this capture's Claim 1 (accumulated-threshold triggers dominate real systems) bears on directly.