Runtime-Independent Persistent Agents: Preserving Identity, Memory, and Code Across Models, Harnesses, and Servers
cs.SE · cs.AI
Quick verdict
Proposes I/M/B agent-boundary + migration protocol, implemented in Enoch; honest but small supervised demos support mechanical continuity only.
Summary
This paper defines runtime-independent persistent agents: the continuity-bearing substrate is a triple of installed identity, private durable memory, and a versioned executable body (I,M,B), separated from replaceable models, harnesses, hosts, and chat surfaces. Six invariants and a quiesce–checkpoint–validate–bind–rehydrate–resume protocol govern authorized migration; the Enoch implementation adds separate body/identity loading, durable private state, five provider contracts, and epoch-based fencing. Evidence spans three frozen snapshots: an operational case that preserved identity and memory hashes, body revision, and 22 historical task IDs across host migration, a Codex-to-Claude swap, and a Slack-to-Telegram swap; a cross-host task resumed from a hash-verified checkpoint; and a supervised Codex–Muse–Codex study where the qualification run, five round trips, and five matched controls all passed on first attempts (28 continuity checks per round trip), plus a planned SIGKILL interruption recovered via the native pause API with stale-reply rejection. All evidence is author-curated and operator-supervised, tasks are trivial fixed-oracle arithmetic, and the authors explicitly do not claim behavioral equivalence, statistical reliability, or portability to arbitrary workloads.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The persistent agent is the continuity-bearing substrate (I,M,B) of identity, memory, and versioned executable body; models, harnesses, hosts, and chat surfaces are replaceable execution/interaction bindings.
- Six continuity invariants plus a six-phase quiesce–checkpoint–validate–bind–rehydrate–resume protocol define authorized migration with a single promotion point, distinct from restart, replication, and fork.
- Enoch preserved an established instance's identity, nonempty memory, body revision, and 22 historical task IDs across real host, joint Codex-to-Claude model/harness, and Slack-to-Telegram substitutions.
- In a supervised frozen study, qualification, five Codex–Muse–Codex round trips, and five matched controls passed on first attempts, and a planned SIGKILL interruption recovered via the native pause API while rejecting a stale reply.
- These results establish mechanical continuity only for the tested deployments and bounded workflows — not behavioral equivalence, arbitrary-task portability, or exactly-once external effects.
Strengths
Unusually disciplined claim scoping
The abstract and Section 9 enumerate exactly what is not established (behavioral equivalence, arbitrary-task portability, exactly-once external effects, unattended reliability); Table 4 maps every invariant to concrete evidence AND an explicit coverage boundary (e.g., I2: exception-rollback tests do not establish crash-atomic conversion).
Reproducibility scaffolding
Three frozen snapshots with hashes (Table 2: B0 c8013ed2, B1 e40b2878, B2 66781e20); B1's original and anonymized distributions each independently passed 963 tests; the B2 Muse archive is Apache-2.0 pinned at commit 084e1b1c6a54 with raw records, migration bundles, and a model-free offline verifier.
Conceptual clarity grounded in distributed systems
Negative-result and process disclosure
I6 is a subtle, practical invariant
Weaknesses
Tiny, inference-free evaluation
Fully self-attested evidence
Trivial workloads
Supervision contaminates independence
Novelty boundary rests on an unverified preprint cluster
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper (arXiv 2609.00546v2, cs.SE/cs.AI) asks a lifecycle question: what must persist for two executions at different times to count as the same agent when the model, harness, host, or chat surface changes? Its answer is an architecture, not a behavioral benchmark: a persistent substrate P_t=(I_t,M_t,B_t) — installed identity, private durable memory, versioned executable body — split from replaceable execution E_t=(R_t,H_t,D_t) and interaction surfaces S_t (Section 2, Fig. 1). Six invariants (I1 identity/lineage through I6 self-description) and a six-phase quiesce–checkpoint–validate–bind–rehydrate–resume protocol (Section 4, Fig. 2) define authorized migration with a single promotion point. Table 1's operation taxonomy (restart/resume/migrate/evolve/identity-update/replicate/fork) is genuinely clarifying. Claims are deliberately scoped to mechanical continuity for tested deployments and bounded workflows, with behavioral equivalence, arbitrary-task portability, exactly-once external effects, and unattended reliability explicitly excluded.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is a runtime-independent persistent agent?
An AI agent defined by what survives replacement of its execution environment: an installed identity, private durable memory, and a versioned code 'body' (together I,M,B). The paper argues the model, agent harness, server, and chat app are replaceable bindings, and that moving between them should preserve lineage and a single continuation authority rather than silently create a new agent or duplicate authority.
Does the paper show an AI agent keeps its personality after switching models?
No. The authors explicitly do not establish behavioral equivalence — only 'mechanical continuity': identity records, memory state, code revision, and task lineage survive the swap. Self-description probes were minimal, and behavioral fidelity is deferred to complementary benchmarking work (PAI-Bench), not demonstrated here.
Is the Enoch code and data available?
Partly. The Enoch body and the Muse adapter are open source; the B2 Muse experiment archive is Apache-2.0 and pinned at a specific commit with raw records, migration bundles, and an offline verifier that needs no model calls. However, the operational case's private identity, memory, and conversations are withheld, and live reruns require Codex/Claude/Muse access plus a human operator.
How large were the experiments?
Small and supervised: one operational migration case, one synthetic cross-host two-stage task, and a Codex–Muse–Codex series of one qualification run, five round trips, five matched controls, and one planned SIGKILL fault. All passed on first attempts, with no statistical analysis — a limitation the paper itself discloses.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The referenced ancillary evidence (INVARIANTS.md, redacted operational record, B2 study archive, claim-to-evidence map) was not part of the provided text, so its internal consistency could not be independently checked.
- Model identities (gpt-5.6-sol, claude-opus-5, gpt-6-astra) are the paper's own client-recorded labels per its text; this review cannot verify backend attestation, and the review takes recorded version strings (Codex CLI 0.153.1, Claude Code 2.1.258, Python 3.12.x/3.13.x) at face value.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.