AI can answer.
Can it remember?

Imagine a notebook that gets harder to use as it fills up. Effective Indexed Memory Capacity is the largest tested number of facts where the memory score still reaches 80. It describes this model-and-notebook setup. It is not the model’s context window or the amount of data stored on a disk.

“No tested load passed” does not mean no memory exists. “At least 256” means we reached the largest tested load before finding a limit. All curves and threshold sensitivity are in the data.

The combined score uses a geometric mean. A zero on any of the six checks makes the combined score zero. Open the component table below to see which checks failed.

For each model we also run two opposite histories, A and B, plus a second copy of A. We compare the difference caused by history with the difference between copies that saw the same history.

This is a limited CHIS-style sequence: we record state and behavior over time, while CHI stays uncalculated. Public API identifiers do not let us verify that proprietary weights stayed identical.

Read the publication

The idea: A system should be tested on how it remembers, changes and recovers over time. Comparing it with people requires a measured human reference.

Evidence now: Measurement proposal with a specification and calculation examples. The original release reports no measured CHI for a real model. This limited pilot cannot validate the full scale.

What this pilot can test: Notebook memory quality across four loads; short history effects, persistence and correction; small reasoning, choice, framing and integration probes. Results may support or weaken these narrow hypotheses.

What it cannot establish: Human likeness, a valid CHI ranking, the full 30-test panel, lasting individuality, autonomous thought, feelings or consciousness.

Data still needed: A defined human sample and matched conditions; held-out calibration; repeated full-panel measurements; longer histories; reliability and subgroup checks; tests of whether CHI adds value beyond existing benchmarks.

Related research (background, not proof of CPLOM): longmemeval — Prior tests of updates, time, multi-session reasoning and abstention.; stratmem — Prior tests of memory use beyond factual recall.

Read the publication

The idea: Useful long-term memory needs source records, dates, controlled updates and a way to search competing evidence.

Evidence now: Architecture proposal with small offline code examples. Those examples show control flow, not a complete memory system or measured advantage over alternatives.

What this pilot can test: Whether a basic model-written notebook preserves correct values, sources, older facts and authoritative updates after a reset, and avoids unsupported answers.

What it cannot establish: The proposed hierarchy, retrieval routing, competing evidence contexts, repeated challenge, deletion controls or production scale. The pilot notebook does not implement that architecture.

Data still needed: An implemented memory system; the same answering model and matched budgets across full-context, flat-retrieval and hierarchical alternatives; tests that remove each proposed feature; labeled evidence, update, access and deletion cases; cost and latency logs.

Related research (background, not proof of CPLOM): rag — External retrieval is established prior work.; memgpt — Prior memory tiers and context-management architecture.

Read the publication

The idea: Reliable AI needs a process for checking constraints, uncertainty and alternatives before taking action.

Evidence now: Architecture argument informed by the author's deployment account. Broad reliability and economic claims are not established by a controlled comparison in this release.

What this pilot can test: How the tested model-plus-notebook systems follow supplied constraints, revise choices and respond to missing evidence in small artificial tasks.

What it cannot establish: Whether governance causes safer production decisions, whether shared-agent coordination improves operations, or whether total business costs fall. No governance architecture comparison is run.

Data still needed: Matched systems with and without specified checks; recorded actions and outcomes; defined failure and escalation rules; tests under disruption; workload, staffing, compute and total-cost records. A causal claim needs a credible control design.

Read the publication

The idea: Giving agents opposing roles and a formal approval process may catch errors that agreement alone misses.

Evidence now: Architecture proposal plus author-reported operational observations, not independently reproduced in this release. Reported route counts, correction rates, error reductions and savings lack the underlying operational data and control comparison here.

What this pilot can test: Only separate model behavior on conflicting records and declared source rules. These are memory probes, not a test of QPM.

What it cannot establish: QPM's advantage over voting or other review methods; the reported 37,500 routes with 22 corrections; roughly 20% residual-error reduction; labor savings or safe operation. A low manual-correction rate is not automatically a true error rate.

Data still needed: Route-level input, output and review logs; a clear error definition; independent review including uncorrected routes; baseline denominators and dates; QPM rules and versions; matched-compute voting and challenge controls; staffing, cost and service-outcome records.

Read the publication

The idea: Coordinating routes, warehouses, staffing and computing may reduce unstable swings and extreme delays across the whole system.

Evidence now: Control framework plus author-reported operational observations, not independently reproduced in this release. Reported stability and transfer across deployments need underlying time-series data and a control design.

What this pilot can test: Short changes in model choices after changed evidence, neutral notebook steps and a correction. This is a small behavioral analogy to tracking change, not a logistics stability test.

What it cannot establish: Reduced delivery-time variance, fewer extreme service failures, weaker cross-region cascades, stable production control or transfer to other businesses.

Data still needed: Timestamped state, demand, weather, resource, intervention and service logs; fixed definitions of stability and extreme events; periods before and after rollout; comparison regions or phased rollout; uncertainty estimates and records from each claimed deployment.

Read the publication

The idea: The author reports that coordinated predictive control increased delivery volume while reducing delivery time and active staffing.

Evidence now: Author-reported operational observations, not independently reproduced in this release. The reported 210% volume increase, 36% shorter average delivery time and 17% fewer active couriers are not results of this pilot.

What this pilot can test: No direct deployment claim. Its memory and history tasks can help shape later model tests but contain no deliveries, dispatchers or warehouse operations.

What it cannot establish: The reported throughput, time and workforce changes; dispatcher time savings; market effects; or whether CPLOM caused these changes rather than demand, staffing or other changes.

Data still needed: Original delivery, route, staffing and dispatcher-time records; exact dates and metric definitions; workload and service-quality measures; all concurrent process changes; a comparison group or credible rollout design; full costs and independent recalculation.

Read the publication

The idea: Combining forecasting, operational indices and rule checks across logistics layers may improve throughput and decision reliability.

Evidence now: Technical architecture description plus author-reported operational observations, not independently reproduced in this release. Equations describe the method; they do not verify the reported production effects.

What this pilot can test: No direct production claim. It can measure only the declared notebook and short-history tasks on the selected public model endpoints.

What it cannot establish: The reported 210% volume growth, 36% delivery-time reduction, 17% active-driver reduction, about 99.82% decision accuracy or about 800 ms computation time; production scalability or portability.

Data still needed: A versioned implementation and evaluation dataset; decision labels and denominators; raw timing and load traces; delivery and staffing records; comparison periods and systems; per-deployment results; tests separating forecasting, voting and rule checks. Agreement between repeated runs must be checked against actual correctness.