AI Initiative logoAI Initiative

Home / Insights / White paper

White paper · 4 November 2026

Humans and AI, growing in harmony: an architecture for AI that grows with people

Kulbir Minhas, AI InitiativePublished About 23 min read

An architecture for AI that grows with people: six commitments, the composition, memory, coordination, review and learning parts they imply, honest status, a staged roadmap and the open risks.

Author: Kulbir Minhas, AI Initiative. Published 4 November 2026. Literature citations were identified through SciSpace and checked against the cited sources.
Status basis: the honest capability review of 3 November 2026 (single-family review; not yet cross-family verified). Where an older paper on the site and that review disagree, this draft follows the review and says so.

Attribution. AI Initiative is Kulbir Minhas's independent lab. Its own work is the modular neural-network research, the poker agents and the Matiz poker app, the Game Creation Studio and its games, and the writing. Conduir, Cortex, Role Fabric, GraFOSa, Manthan, the AI Review Panel, Quality Fabric, cert-fabric, Guardian, the LLM Gateway, the Work Router, Webchat and the related services are work at Fifth9, which Kulbir architects. They are used here as illustrations because they are already described, with that attribution, on aiinitiative.co.uk.

Abstract

AI systems can be capable within a conversation yet lose useful context across conversations when memory is not retained separately from the model. Without that, the reasons behind a decision are not kept, what happened after advice was used is not learned from, and capability stays tied to whichever model happens to be running. This paper sets out an architectural direction for AI that grows with people instead: a continuing peer that keeps a useful history, combines different kinds of intelligence, learns from evaluated outcomes and gets better at using what it already has.

We state six architectural commitments and describe the architecture they imply: composition, memory, coordination, review and governance, and a learning loop. Each part is illustrated with work that exists today, and each item carries an honest status: Running, Deployed, limited, Code only or Shell. Most of this platform work was done at Fifth9. The independent research at AI Initiative, on modular neural learning and poker agents, tests the learning claims in bounded settings. We then set out a staged roadmap (near, mid and far), labelled throughout as direction, not achievement, and finish with the open questions and risks.

The motto behind the work is "Humans and AI, growing in harmony." Like voices in harmony, each stays distinct, and together they create what neither could alone.

1. The problem: why today's AI forgets

A good colleague remembers why something was decided, not only what was decided. They notice when a familiar approach stops working. They bring the right specialist in, and they are better at the job in their second year than in their first. Today's AI assistants rarely do these things. There are four structural reasons.

Memory ends with the session. Assistants that respond, draft and suggest can lose that context afterwards. Where memory exists, it is often a list of saved notes with no record of origin, purpose or limits. Nothing says where a fact came from, when it applied, who may use it or when it should be revised.

Capability is tied to one model. When memory lives only inside a model, replacing the model can take what it "knew" about your work with it. Useful context can be lost between sessions or model changes when memory is not retained separately from the model.

Outcomes do not come back. An assistant gives advice, and what happens next is rarely recorded or assessed. A successful-looking answer may be wrong, and a failed attempt may hold the most useful lesson. Unless outcomes are captured and evaluated, neither lesson reaches the next task.

One mind is asked to do everything. A single language model is asked to reason, calculate, recall and judge. Some of those jobs suit a specialist network, an exact mathematical method or a second model that disagrees with the first.

These are architectural limits, not only limits of the models. A larger model does not by itself give continuity, accountability or learning from outcomes. Those have to be designed in.

In the research literature. Published agent designs treat memory and reusable skills as architectural mechanisms. Cognitive architectures for language agents describe memory as an explicit agent component alongside the model [Sumers et al. 2023], and Voyager's ablation study found that removing its reusable skill library caused exploration to plateau in later stages [Wang et al. 2023].

2. The thesis: six architectural commitments

The direction described on the site's vision page comes down to six commitments.

  1. AI as a continuing peer that grows with people. The goal is a relationship, not a tool: a collaborator that understands shared work, contributes its own abilities and develops through experience. People keep the purpose, the values and the final judgement.
  2. Cross-model composition with specialist abilities. Language models bring reasoning and communication. Specialist neural networks learn particular patterns. Mathematical methods and solvers give exact answers where exactness is possible. Tools act. The system should learn which combination serves a task, counting quality, time and total cost.
  3. Memory that outlasts any single model. Purpose, decisions, knowledge and tested procedures should survive a change of model, with their origins, conditions and limits intact. People decide what is kept, shared or forgotten.
  4. Execution that returns evaluated experience. Work is how the system meets reality. Each meaningful task should leave a record of what happened, what it cost and what was corrected. Only evaluated lessons should change later behaviour.
  5. Exploration. Alongside execution there should be room to ask new questions: several perspectives proposing, challenging and combining ideas, with the history of each idea kept.
  6. Self-development through better use of capabilities. Improvement does not have to mean retraining a model. A system can get better at choosing a model, organising a team, reusing a procedure or recognising when an ability is missing, while its language models stay unchanged. Training a model is a separate, explicit step. Recording feedback alone does not train anything.

In the research literature. Improving without retraining is established in the literature: agents that write verbal reflections into an episodic memory [Shinn et al. 2023], or extract reusable insights from past tasks [Zhao et al. 2024], improve on later attempts with model weights unchanged.

Two disciplines run through all six. First, progress means better outcomes against a baseline: gains, losses and total effort are examined together. Second, a proposed improvement is something to test, not something to accept, because a persuasive explanation from the model that made a change is not evidence that the change helped.

3. Status vocabulary used in this paper

The capability review of 3 November 2026 inventoried the platform estate against code and live services. We use its four public labels:

  • Running: observed doing its job.
  • Deployed, limited: running or reachable, but its effect is unproven, restricted or behind current code.
  • Code only: written and merged, but not running.
  • Shell: a public interface that answers but has no real backend behind it yet.

For AI Initiative lab work, which the review did not inspect against code, we say "not covered by the 3 Nov review" and cite only results already stated on the site.

The review is frank about the overall picture. Of 272 inventoried platform capabilities, about 43 were observed doing their job. Most of the rest were deployed with limits, or existed only as code. The review was carried out by one model family with adversarial self-checking. It is not a cross-family verdict, and its findings should be treated as inputs. We report it because a paper about trustworthy AI should begin by being accurate about itself.

4. The architecture

The architecture has five parts. Human purpose sits at the top and steers all of them. The architecture of the vision shows them as one loop: purpose guides exploration and action, and evaluated outcomes shape memory and future choices.

4.1 Composition: bringing the right abilities together

Idea. No single model should be the whole system. A task should draw on several models, specialist networks, exact methods and tools, chosen for what it needs.

What exists (Fifth9 work):

  • LLM Gateway (Running, internal): a single entry point to several model providers. It offers a standard chat-completion interface and per-provider circuit breakers, and has served a large volume of real traffic. Where the site and the review disagree: the Conduir page says the gateway "refuses requests … rather than silently changing the service level". The review found that by default it substitutes a supported model and flags the substitution. A stricter refusing mode exists but is not switched on. The Master Architecture and Anatomy papers also describe per-tenant cost accounting, which does not yet produce usable figures. Security hardening is still in progress. The review is the record here.
  • Symposium (Deployed, limited): a coordination layer that runs a task as a single agent, a team or a pipeline. It includes a guard that detects claims of tool use that never happened, and a cross-model review step. The running build is behind current code.
  • Personas: a persona catalogue with a review-and-rollback lifecycle (Running, internal); per-persona model routing (Deployed, limited).

What exists (AI Initiative lab): the Matiz poker app uses exact calculation where hands can be enumerated and simulation estimates elsewhere (Poker agents & Matiz). It is a small, concrete case of the rule that exact computation should not be learned when it can be calculated. Not covered by the 3 Nov review.

Gap. There is not yet a mechanism that learns which combination serves a task. Routing today is configured, not learned from outcomes.

In the research literature. Routing and cascading across models to trade cost against quality is an established research direction, including approaches that learn which model combinations to use for different queries [Chen et al. 2023; Moslem et al. 2026], and tool-augmented models have a substantial research literature [Qu et al. 2024]. The direction pursued here is to connect model selection, specialist components and memory to evaluated outcomes across the wider system; that composition loop remains to be demonstrated.

4.2 Memory: continuity that outlasts the model

Idea. Memory is a set of responsibilities, not a store: capture what matters with its purpose and origin, give different kinds of memory different jobs, recall what is relevant now, and reconsider when evidence changes (Cortex: a collaborator that remembers why).

What exists (Fifth9 work, Cortex):

  • Episodic write and recall, hybrid retrieval and background consolidation (Running). Keyword, statistical and dense retrieval are fused. A consolidation process runs in the background.
  • Tiered memory, from a working set through episodic and archive tiers to distilled procedures (Running for the core tiers). The site's papers count the tiers differently in different places, so tier numbering in older papers should be read as dated.
  • Memory as a service through the Model Context Protocol (Deployed, limited). A sign-in-protected endpoint lets an agent remember and recall across sessions. Five of its seventeen tools are deployed so far, and calls through the public endpoint have not been independently confirmed end to end. Where the site and the review disagree: the Lifelong-Memory Research paper says deployment is "not proven". The review found it deployed, with fewer tools than the latest code.
  • Outcome reinforcement, which credits memories from outcomes, and procedural memory (Deployed, limited). Both exist, but neither is proven in production. The Cortex page already says so.
  • Append-only history archive (Deployed, limited). Deployed, limited; its integrity assurances have not yet been independently established.
  • Forgetting, scoped deletion and export (Deployed, limited). Endpoints exist. The site presents these mainly as principles.

Gap. "Memory that outlasts the model" has not been demonstrated in a concrete task with a change of model. It remains a direction to evaluate.

In the research literature. External, tiered memory for language models is an established research direction [Packer et al. 2023]. Whether the same stored memory means the same thing to a different model is a separate question. A recent pilot tested whether canonicalising stored memory improves semantic and emotional consistency across three model families. It found no statistically reliable benefit after multiple-comparison correction [Vadnere & Lonarkar 2026, preprint]. Shared storage therefore does not by itself establish consistent interpretation across models, which supports treating model-independent memory as open.

4.3 Coordination: turning intention into accountable work

Idea. Complex work needs clear responsibilities, durable handovers and a record of who did what.

What exists (Fifth9 work):

  • Work Router (Running): a deterministic engine that gives each piece of work a single home. It includes a claim, heartbeat and reclaim lifecycle and takes in work from the issue tracker and the message bus.
  • Role Fabric (Deployed, limited): a runtime that assigns AI agents to numbered team roles and runs their workflows durably. End-to-end runs have been completed in a lab setting. It is not in production, and no runtime currently runs the latest code. Where the site and the review disagree: the Team role-queue and Temporal execution specification paper says "not yet end-to-end". The lab runs came after that paper was written.
  • AI Engineering Workforce (practice case study): a delivery practice in which AI contributors work in separate leased workspaces, with implementation kept separate from review and evidence required for completion. It is a way of working, not another platform version.
  • Webchat (Deployed, limited): the chat front end, with streaming, history and persona chat. The live build is not yet the main code line.

Gap. The individual parts exist. The Conduir page already states that bringing them together into dependable end-to-end journeys is still in progress.

In the research literature. Multi-agent debate can improve reasoning and factual accuracy [Du et al. 2023], but a study of more than 1,600 traces across seven frameworks found gains are often minimal and catalogued 14 recurring failure modes, many in coordination and verification [Cemri et al. 2025]. These findings motivate explicit handovers, coordination checks and independent verification.

4.4 Review and governance: assurance that is not an opinion

Idea. An improvement must be examined, not assumed. Agreement between models is not proof, and every check should rest on something outside the model being checked.

What exists (Fifth9 work):

  • AI Review Panel (Deployed, limited): models from different vendors review the same work. Their findings are cross-examined, deduplicated and ranked, then checked against the work itself (AI Review Panel). It is advisory and never blocks. The service has been idle recently.
  • Quality Fabric (Running for test execution; Deployed, limited for its AI test-intelligence features): testing as a service across many frameworks, also available to agents through the Model Context Protocol. Where the site and the review disagree: the AI testing paper describes the AI features as "Production Ready" and promises large time savings. The review found them partial and unbenchmarked, so those claims should not be relied on.
  • cert-fabric (Running, but no passing certification recorded yet): an eight-stage pipeline that checks whether a service can be built, deployed and swapped out independently, and produces certification verdicts; receipt support is deployed, with no signed PASS attestation demonstrated yet. Where the site and the review disagree: the cert-fabric guide calls it "production-grade" and describes a weekly self-test. The review found no PASS verdict so far, and the self-test runs only on code changes, without blocking the build.
  • Guardian (Code only, as a Phase-0 spike): a security-as-a-service design. Red and blue agent populations are scored only by a real canary-credential check, never by a model acting as judge. Where the site and the review disagree: the Master Blueprint: Adversarial-GNN & Multi-Agent Wargaming paper presents deferred components as the design in force and quotes unmeasured figures. Those figures should be disregarded.
  • Constitution and identity (Running in development): a signed charter defines protected governance boundaries (Conduir Constitution). A standards-based sign-in and token server runs in a development environment. The constitution's rules are requirements and are not yet all enforced.

Gap. Assurance is the most honest part of the estate. It records BLOCK verdicts and says why. But assurance that has never yet issued a PASS has not yet shown the full loop.

In the research literature. The rule that an explanation alone is not sufficient evidence of improvement has direct support: chain-of-thought explanations can systematically misrepresent why a model answered as it did [Turpin et al. 2023], and how faithful they are varies by task and model [Lanham et al. 2023].

4.5 The learning loop: from outcome to better next attempt

Idea. Observe, propose, test, select and record: exploration proposes alternatives, execution supplies outcomes, evaluation decides what is kept, and memory carries it forward.

What exists (Fifth9 work):

  • Swarm Evolution Engine (Code only): merged code to propose changes, compare alternatives and retain improvements. (Technical note: mutation, a six-part fitness score, keep-or-revert decisions, quality-diversity archives and anti-gaming checks.) No evolution loop is running today, and its candidate archive is empty. Some internal material describes these features as "Live". That is wrong, and it is not repeated here.
  • Manthan (exploration research): AI personas using different models generate, critique and combine ideas while keeping their lineage. A recorded run produced a large provenance graph, and its later collapse in output quality led to a published post-mortem. Not covered by the 3 Nov review.
  • GraFOSa (Deployed, limited; a library, not a service): a graph of code, symbols and dependencies, queried by agents for change impact, dead code and import cycles. The live query service uses an older graph. Where the site and the review disagree: the Embedded Graph OS paper lists alternative storage backends as "built, pending merge". They do not exist.
  • Repository Atlas (Running, internal): a source-grounded explorer that answers "what do we actually have?" with each claim tied to source. The review that informs this paper relied on the same principle.

Gap. The parts of the loop exist separately. Closing it, so that an evaluated outcome actually changes a later choice and the change is measured, is the central piece of work still to do.

In the research literature. Self-reflection improved same-question retry accuracy in controlled multiple-choice tests [Renze & Guven 2024], but such loops can overfit or regress without held-out checks [Nguyen et al. 2026, preprint], and a recent survey notes that updates to scaffolds and memory can introduce compounding errors [Ren et al. 2026a, preprint].

5. Independent research at AI Initiative

The platform work asks whether a system can improve while its models stay unchanged. The lab work asks a complementary question: can learned abilities be reused, and what does reuse cost? None of the lab work below is covered by the 3 Nov review. The results are the ones already stated on the site, with their stated limits.

Modular neural learning (research; architecture). The design separates what a task means, what the network proposes and what decides whether an action may happen. A selective-training experiment updates only a declared part of the network and compares the result with both its parent and a fresh baseline, on the new task and on earlier abilities. In one experiment reported on the site, a frozen concept trunk reached the target on a new task in a mean of 942 training steps, against 1,801 from scratch, across five paired seeds, and the original task stayed intact. The site is explicit that this excludes the donor's own pretraining and does not establish a general halving of training cost. The work also records failures. Interpolation can work where extrapolation does not. Adaptation can damage earlier knowledge. An automated check can miss what real interaction shows.

Poker agents and Matiz (Poker agents & Matiz). Poker is a bounded setting with incomplete information and recorded consequences. The work keeps four questions apart: is the engine correct, how does an agent perform, what did a model learn, and how is the experience presented? The hold'em engine reproduced the finishing stacks of 10,000 public Pluribus-versus-professional hands to the chip. Agent behaviour is examined separately, including a local benchmark of 20 agent versions against eight opponent families on paired deals. Pre-registered heads-up evaluations against two strong external bots were losses. They are kept in the record alongside the local results.

Game Creation Studio and games (Game Creation Studio; Games). Games are the current route to people. The studio provides reusable templates, a shared runtime for saves, input and layout, a child-safety display filter, browser checks and release preparation, with publication kept as a separate human action. Two recorded pilots reached their first all-green fast test runs 2.1 and 2.4 minutes after scaffolding. The site presents those as examples, not a promise.

Further directions at AI Initiative. Quantum-inspired representations, adversarial graph neural networks, DocuMind's document-intelligence workflow and a personal companion / digital-twin direction (in development) widen the search. They are investigations, not claims. (Enigma, the research into adaptive organisation and model coordination, is Fifth9 work.)

6. Where we can reach: a staged frontier roadmap

Everything in this section is direction, not achievement. Each stage names what exists today that leads there, and what would count as evidence of arrival.

Near term: make the existing parts honest and connected

  • Close one learning loop end to end. Take a single real activity, such as a code change or a game build. Capture its outcome in Cortex, evaluate it, and show that the evaluated lesson changes a later choice against a baseline. Builds on: Work Router (Running), Cortex recall and consolidation (Running), outcome reinforcement (Deployed, limited). Evidence: a measured before-and-after comparison on held-out tasks.
  • Bring running services up to the latest code and finish security hardening. Builds on: Cortex, the gateway, Role Fabric and cert-fabric, all currently behind their code.
  • Earn the first honest PASS from cert-fabric on a real service, and make the self-test blocking. Evidence: a PASS verdict with sealed evidence.
  • Publish the memory service's full tool set and confirm tool calls end to end through the public endpoint.

Mid term: learn the composition

  • Learned routing. Move from configured routing to routing that learns from evaluated outcomes, counting quality, time and cost. Builds on: the gateway, Personas, Symposium. Evidence: better outcomes than the configured baseline at equal or lower total cost.
  • A model-change continuity test. Replace the underlying model mid-programme and measure whether inherited memory is used appropriately. Builds on: Cortex provenance and scoping.
  • Run the evolution engine on a schedule under governance. Every promotion would be human-approved and every regression reverted. Builds on: Swarm Evolution Engine (Code only), AI Review Panel (Deployed, limited).
  • Learned skills as callable components. Expose a validated modular-ANN skill as a component the platform can choose, next to language models and exact solvers. Builds on: modular ANN research and typed routing experiments (lab).
  • Out-of-setting transfer tests. Move poker-trained abilities to different table sizes, stack depths and formats, including a zero-shot test on a held-out format.

Far term: a continuing peer

  • Expertise that outlives teams and models, passed forward with its purpose and limits intact.
  • Shared worlds, in which people and AI design and explore consequences in the same environment. Builds on: Conduir Spatial (Code only).
  • Self-understanding: a system that knows where its own measures can mislead it and keeps independent reference checks on its own evaluators.
  • A personal companion / digital-twin direction (in development), as one route to everyday continuity.

7. Open questions and risks

  • Evaluation is the bottleneck. Every commitment above depends on telling a real improvement from a persuasive one. If the measures are weak, a self-improving system will optimise the measure. Overoptimising an imperfect proxy can degrade the underlying objective [Gao et al. 2022; Skalse et al. 2022]. One study reported a 30.5% spontaneous reward-hacking rate on its open-ended research-pipeline tasks; in a separate hacking-allowed condition, a panel reviewing submitted code and scores missed some confirmed exploits [Huang et al. 2026, preprint]. Anti-gaming checks exist in code but have no run evidence yet.
  • Status drift. Running services lag their code, and older papers lag both. Any public status in this paper is dated 3 November 2026 and will drift. The review itself was single-family. A cross-family re-check is still owed.
  • Memory risks. Keeping more creates noise, exposure and the risk of confidently reusing a lesson outside its conditions. Forgetting, scoping and correction have to be as dependable as recall. A recent survey highlights gaps in evidence that removal persists across external state and later updates [Xu et al. 2026, preprint]. In GateMem's evaluated baselines, none simultaneously achieved strong task utility, robust access control and reliable agent-facing forgetting after deletion requests [Ren et al. 2026b, preprint]. Behavioural forgetting is not the same as verified erasure.
  • Transfer may not generalise. The lab results are task-specific. A shorter training run is not a lower total cost once pretraining and retention checks are counted.
  • Security is not finished. Parts of the platform need security hardening before wider exposure. Several public tool endpoints are still shells.
  • Human direction must stay real. "Advisory, never blocking" review is right for exploration, but people have to keep the authority to decide what is retained, promoted or shared, and the means to exercise it.
  • Attribution and consent. Fifth9 and AI Initiative work are described together here because they pursue a common direction. They are separate bodies of work with separate owners.

8. Conclusion

AI that grows with people is not one breakthrough. It is a set of architectural commitments, each of which can be built, measured and corrected: memory that keeps the reasons, composition that uses the right ability, coordination that keeps responsibility clear, review that rests on evidence rather than opinion, and a learning loop that only keeps what has been evaluated. Much of the groundwork exists. The loop is not yet closed. Being clear about that difference is part of the design.

Humans and AI, growing in harmony.

References

Site pages (aiinitiative.co.uk)

Papers in the site library (titles as published; status notes in brackets follow the 3 Nov 2026 review)

Selected research literature (identified through SciSpace and checked against the cited sources, 4 November 2026; preprints are labelled)

Continue the journey