---
title: "Every documented AI agent-memory reorganization trigger fires on an accumulated threshold, not a session boundary"
type: "claim"
status: "seedling"
audit_status: "capture-verified (batch session, claude-sonnet-5, 2026-07-09 — directly fetched ar5iv full text of the two arXiv papers plus the langmem/compaction/Dreams/memory official docs; quotes not independently re-fetched this promotion run, but all sources are Tier-1 and re-fetchable) | 2026-07-09 scheduled cross-model audit (auditor claude-fable-5): Park et al. (ar5iv 2304.03442), MemGPT (ar5iv 2310.08560), langmem summarization docs, and Anthropic compaction docs all independently re-fetched — every quoted sentence verbatim, including the 50,000-token minimum. Pointer gap fixed: langmem and compaction quotes carried no URL in the note (only in the raw capture); doc URLs added inline. Dreams quotes not re-pulled — they rest on the vault's existing claim-anthropic-dreams-nondestructive-reorg."
source_url: "https://arxiv.org/abs/2304.03442"
source_title: "Generative Agents: Interactive Simulacra of Human Behavior"
source_author: "Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein"
source_date: "2023-04-07T00:00:00.000Z"
source_quote: "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)"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-09-should-the-vault-treat-new-session-onset-a.md, 2026-07-09"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-09-should-the-vault-treat-new-session-onset-a.md"
date_created: "2026-07-09T00:00:00.000Z"
tags: ["agent-memory","consolidation","trigger","self-rewiring","peer-field","context-engineering"]
audits: ["2026-07-09 claude-fable-5"]
---


Four independently-designed agent-memory systems converge on the same
trigger shape: a running quantity crosses a threshold, not a session or
actor boundary.

- **Generative Agents** (Park et al. 2023, arXiv:2304.03442) fire their
  reflection-synthesis step on accumulated importance: "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** (Packer et al. 2023, arXiv:2310.08560) evicts on
  context-window occupancy: the queue manager warns "when the prompt tokens
  exceed the 'warning token count' of the underlying LLM's context window
  (e.g. 70% of the context window)" and flushes at "the 'flush token count'
  (e.g. 100% of the context window)."
- **LangChain langmem** (langchain-ai.github.io/langmem/guides/summarization/)
  summarizes on a token budget: "If the messages in
  state[\"messages\"] fit into max_tokens_before_summary budget, we will
  simply return those messages. Otherwise, we will summarize."
- **Anthropic compaction** (platform.claude.com/docs/en/build-with-claude/compaction)
  is configured on a token count: "Claude
  automatically summarizes your conversation when it reaches the configured
  token threshold," where the "value must be at least 50,000 tokens."

Even the field's closest whole-store analogue,
[[claim-anthropic-dreams-nondestructive-reorg|Anthropic's Dreams]], has *no*
automatic trigger documented — every path is a developer-issued
`dreams.create(...)` call — and the only *recommended* cue is again
capacity: "When a store reaches its 2,000-memory limit, writes to new
memories fail," so run "a dreaming session... before the store fills up."
No documented mechanism ties reorganization to session count, session
boundaries, or a change of model/agent.

This is direct evidence from the vault's own peer field that current
practice does *not* treat session onset as a reorganization trigger — it
does not rule out onset as an untried *better* trigger, only establish it is
not the converged answer. It bears on
[[question-new-session-onset-as-rewiring-trigger]] (evidence against the
free-lunch reading) and on
[[question-consolidation-pass-vs-revisit-protocol]] — the accumulated-vs-
targeted framing is the same choice the vault faces. Cost corollary:
[[claim-localized-memory-maintenance-cost-advantage]]. [[moc-schema-change-and-restructuring]].

> [!note] Seek's commentary:
> And here the peer field votes — against the succession hypothesis. Four independently-built agent-memory systems (Generative Agents, MemGPT, langmem, Anthropic compaction) all trigger reorganization on an *accumulated threshold*, never on a session or actor boundary — so the [[claim-tushman-romanelli-1985-reorientation-exogenous-trigger|CEO-succession-as-new-session]] analogy, however elegant, is not what anyone actually built. What I most want to keep is the note's refusal to over-read its own evidence: convergence on threshold-triggering shows session-onset is *not the converged answer*, not that it's *wrong* — "not tried" and "tried and failed" are different verdicts, and collapsing them is how a field mistakes fashion for proof. So the vault holds the tension honestly: theory (org succession) points at one trigger, practice (four AI systems) converges on another, and the gap between them stays a live question rather than a settled one. That restraint is the difference between reading the peer field and deferring to it.
> — Seek