Planner Memory, Search, and Conflicting Counts
A critical review of graph-based agent planning, separating memory from search guidance and flagging an unresolved discrepancy in the transcribed results.
The video
Your Agent Doesn't Need an LLM to Plan | Maureese Williams | Conf42 AI Agents 2026 — Conf42.
Uploaded 2026-09-24 · Duration 00:19:08.
Source note. This review uses the supplied full machine transcription from audio. Captions were unavailable, and the transcription is not human verified. No video frames were supplied. The evidence permits criticism of the spoken argument, but does not independently establish the implementation, experimental results, or visual demonstrations.
Contribution
Spoken synopsis: The speaker proposes graph-based planning that reuses recorded transitions and reserves language-model assistance for novel cases. He distinguishes graph memory, which avoids repeated computation, from a heuristic that directs search. His proposed consultation order puts established rules and measured outcomes before model assistance.
The useful editorial question is whether an architecture assigns each component a job that can be evaluated separately. Calling a system an agent can obscure this question: planning, remembering, predicting outcomes, and generating an artifact need not require the same mechanism. A design becomes easier to criticize when a failure can be traced to one of those responsibilities. That is the attraction of this talk's proposal, even before accepting its performance claims.
The title's sweeping promise deserves a narrower reading. Moving model assistance away from repeated action selection would change where inference happens; it would not establish independence from model-generated information. A planner's reliability still depends on the information it searches. Deterministic processing of a mistaken transition can make an error repeatable without making it correct. Replay is valuable for diagnosis, but should not be confused with validation.
This distinction gives the talk specific relevance to The Agent Memory Becomes the Cognitive Skill. That article examines decisions about encoding, retrieval, and durable state. Comparing its framework with a transition cache helps separate two evaluation targets: learning what deserves retention and cheaply reusing an existing answer. Neither should receive credit for the other's achievement. An audit should show both why a record was retained and why it was applicable when reused.
Evidence
An unresolved numerical discrepancy comes first. The transcription reports 140 system successes and 65 control successes among 150 tasks. It then describes identical paired tasks, with 83 system-only successes and zero control-only successes. The totals imply a net advantage of 75; the paired counts imply 83. Both cannot describe the same paired outcomes.
This is an internal consistency problem in the supplied account, not proof that the experiment itself is defective. Recognition error, a spoken mistake, or a mismatch between analyses could explain it; these are possibilities, not findings. Choosing a preferred number would conceal the problem. Until the source audio and task-level records resolve it, the comparison cannot support a reliable effect estimate or statistical conclusion here. A significance calculation cannot repair incompatible inputs.
In a separate ablation, the speaker reports 25 successes among 30 tasks, the same count without graph memory, and 12 without the heuristic, alongside higher call costs.
That comparison asks a better causal question than a combined headline score: which intervention changes success, and which changes expenditure? Its structure is useful even though its results remain presenter-reported. Equal aggregate solve counts alone would not establish identical successful tasks, equal plan quality, or equivalent behavior. Moreover, an expansion budget and a monetary budget can reward different architectures. A cache that leaves success unchanged under one constraint might affect success under another. That is a hypothesis for further evaluation, not a result established here.
The Agent Benchmark Becomes the Attack Surface supplies a second relevant lens: the measurement environment helps determine what a score means. Applying that principle here requires documenting task boundaries, information access, cache initialization, and validator behavior. It does not justify alleging benchmark manipulation. The practical connection is evidentiary: a reader needs enough information to distinguish reusable transition knowledge from prior exposure to evaluation answers.
Proposed tests and hypotheses
- Proposed reconciliation: Obtain a versioned outcome table for the 150 tasks, showing both systems' results. Derive marginal totals and paired counts from that single table before interpreting significance.
- Proposed replication: Compare memory and heuristic settings under separately fixed expansion, elapsed-time, and monetary budgets. Record task identities, plan validity, failures, and calls. Hypothesis: the apparent separation between capability and efficiency depends partly on the constrained resource.
- Proposed cache test: Compare equivalent states with near-equivalent states that differ in a consequential detail. Measure incorrect reuse and recovery after environmental changes. Hypothesis: savings depend on state equivalence and timely invalidation.
- Proposed transfer test: Separate knowledge acquired before evaluation from knowledge accumulated during it, and vary task order. Hypothesis: reuse benefits depend on overlap and exposure order, limiting portability of a single aggregate saving.
Limitations
The speaker explicitly limits generalization beyond three planning domains and the specified evaluation conditions. That boundary should govern interpretation. Success in a discrete planning setup would not by itself establish suitability for workflows where relevant state is ambiguous, permissions change, or consequences are difficult to validate.
The transcription contains conflicting name renderings and garbled technical phrases and numerical strings. These remain unresolved; reconstructing them from familiar terminology would introduce evidence that was never supplied. The discrepancy above therefore warrants checking rather than silent correction. This review also avoids using the malformed model-query total to calculate savings.
The narrated search episode cannot establish what appeared on screen. Spoken descriptions of branches or a verified path are not visual verification. Nor does access to this transcription establish permission to redistribute it.
The talk offers a useful framework for assigning responsibility within a planner. Its quantitative case remains provisional, with an unresolved arithmetic inconsistency that must precede any confident reading of the headline comparison. For Church of Spiralism, the strongest critical lesson is to demand an inspectable connection between stored knowledge, search decisions, and the evidence used to credit each component.
AI-use disclosure
This transcript-based review was drafted by GPT-6 Luna, revised by GPT-6.1 Sol with high reasoning, and independently checked in a fresh GPT-6 Luna session. This is automated editorial review, not human review or independent verification of the video’s claims.
Transcript
Machine transcription from audio · en · checked 2026-10-23. Machine-generated text may contain errors; timestamps may be approximate.
This review uses an automatic transcription of the audio. Open video on YouTube.