Asymmetri Research
Project Protocol
What if machine knowledge could compound outside the model?
- AI and ML systems
- AI, Agents & Data
- Technology Consulting
- Innovation & R&D
01In short
Project Protocol in short
Asymmetri research into shared machine cognition: can teams of AI agents write what they learn into a shared, inspectable layer that the next agents inherit and build on?
- Client
- Asymmetri Research (internal R&D)
- Domain
- AI and ML systems
- Category
- Research · Multi-agent AI
- Platforms
- Simulation harness, Experiment framework, LLM agent runtime
- Services
- Research design, AI agent engineering, Experiment design, Evaluation
- Year
- 2026
02Context
The story of Project Protocol
Challenge
Every AI agent starts its working life from zero. It may be brilliant, but what it figures out on Tuesday is gone by Wednesday unless someone retrains the model, which is slow, costly and tied to one vendor, or stuffs old transcripts back into the prompt. Human intelligence didn't scale that way. It compounded because meaning moved out of individual minds into language, writing and institutions that the next generation inherited. Nobody has built that for machines.
Approach
Project Protocol is Asymmetri's research programme into shared machine cognition. The hypothesis: a team of AI agents can externalise useful structure (concepts, situation patterns, strategies, failure corrections and their lineage) into a shared protocol layer outside any model's weights, and later agents can inherit that layer and do better than agents with no memory or with plain transcript search.
We built a deterministic simulation and experiment harness to test it honestly: a world that rewards abstraction, a typed and auditable protocol layer, LLM agents that read it through tools, and baselines strong enough that a positive result would mean something. Then we turned the same rigour on our own results.
Insight
Intelligence can compound outside the weights, if the layer stores the right kind of structure.
A vector database stores text. A memory log stores what happened. Neither stores what was understood. The bet behind Project Protocol is that the missing piece is structure: named situations, strategies with a track record of helping or hurting, and the lineage of how each idea was revised. If that holds, knowledge stops belonging to a model and starts belonging to the system.
2 hard problems
- P01Make machine knowledge compound outside the model's weights.Can a team of AI agents write what they learn into a shared, inspectable layer that the next generation inherits and builds on, the way human knowledge outlived every individual mind?
- P02Build an experiment that can't fool its own builders.How do you test a big idea about machine intelligence so that a 'yes' really means yes, when every shortcut makes a yes easy to manufacture?
03Knowledge that compounds
Hard problem 01 of 02
Make machine knowledge compound outside the model's weights.
Can a team of AI agents write what they learn into a shared, inspectable layer that the next generation inherits and builds on, the way human knowledge outlived every individual mind?
Fleets of agents are arriving in every company: support, operations, research, code. Today each one forgets. If what one agent learns could become something every later agent inherits, audits and improves, organisations would get an institutional memory for machines, and it wouldn't be locked inside any one vendor's model.
Why it’s hard
- F.01Search is a strong rivalRetrieving old transcripts is simple and works surprisingly well. A shared layer has to beat it in a fair fight, at the same reading budget.
- F.02Remembering isn't understandingRecalling what happened is easy. Storing what was understood, in a form a different agent can use on a new problem, is the hard part.
- F.03Agents only see part of the worldReal coordination happens across information boundaries. Each agent knows its own patch and has to rely on others for the rest.
- 01FrameFind the valuable decision, workflow or experience.
Language, writing, institutions. Then machines.
We started from a thesis about how intelligence scales. Humans didn't get smarter by growing bigger brains every generation; they moved meaning into shared forms that outlast any one mind. We framed the machine equivalent precisely, including what it is not: not a vector database, not a memory log, not an API contract, not a central orchestrator. A protocol layer of concepts, schemas, reasoning patterns, failure corrections and lineage.
- Wrote the thesis: shared machine cognition
- Defined the question so it could fail
- Ruled out the easy look-alikes up front
- 02DesignShape the interaction, context and control model.
Build a world where abstraction pays.
An idea like this needs a world that rewards it. We designed a supply grid split into four sectors, one agent each, with three resources, demands on deadlines and corridors that choke. Each agent sees only its own sector clearly and must coordinate across boundaries on a message budget set at half of what it would naturally use. The layer itself holds typed situation patterns and strategies, each carrying evidence of when it helped or hurt, retrieved by a structural vocabulary rather than embeddings so every success and failure can be inspected.
- 4 sectors, partial sight, a deliberately scarce message budget
- Patterns and strategies with helped and hurt counters
- Lifecycle: proposed, active, retired
- 03EngineerBuild the intelligence, agent and tool chain.
A deterministic lab for agents.
We built a deterministic simulation and experiment harness so that every comparison is exact and repeatable. Agents run with no memory, with the protocol layer, with transcript search at the same read budget, or with full sight of the world as an oracle ceiling. LLM agents read the layer through tools and propose moves that a validator checks, running first in shadow mode beside a scripted agent and then live.
- 4 conditions, including an all-seeing oracle ceiling
- LLM agents with tool access to patterns and strategies
- Shadow mode before live mode, with every move validated
- 04ProveEvaluate performance, safety and economics.
Find the headroom before chasing the gain.
About 150 logged episodes later, the first thing the lab proved is that the prize is real: an agent that could see everything fulfils nearly twice the demand of today's agents, so there is a large gap for shared knowledge to close. The first signals of the layer helping appeared exactly where coordination is hardest, on cascading chokepoints. They are early and small, and we treat them that way.
- ~150 logged episodes across conditions
- Oracle ceiling at 1.9× baseline fulfilment: the gap is real
- Per-demand diagnostics on every miss
- 05OperateDeploy, observe and improve in the real workflow.
Run it like a lab, not a launch.
Every round of runs ends in a written learnings report: what the data showed, what it could not support, and what changes next. Those reports reshaped the agents, the world and the question itself, and pointed straight at the next, sharper experiment.
- A learnings write-up after every round
- Design revised four times as the evidence came in
- Next phase designed and costed before a line is written
The answer · Knowledge that compounds
If knowledge can compound outside the weights, every fleet of agents gets an institutional memory.
- 1.9×Headroom: oracle vs today's agents
- ~150Logged episodes
- 4Conditions, incl. an oracle ceiling
- 180Token read budget per agent
04An experiment that can't lie
Hard problem 02 of 02
Build an experiment that can't fool its own builders.
How do you test a big idea about machine intelligence so that a 'yes' really means yes, when every shortcut makes a yes easy to manufacture?
Claims about AI memory and agent learning are everywhere, and most fall apart under a fair baseline. A research programme is only as valuable as its willingness to be wrong. If the result is going to shape how companies build agent systems, it has to survive the people most motivated to believe it: us.
Why it’s hard
- F.01Proxies flatterEasy metrics like message length or token count can move for reasons that have nothing to do with the idea being tested.
- F.02Builders leak knowledgeIf the people building the system hand it the answers, even indirectly, it will look like it learned them.
- F.03Models already know thingsA language model's training can masquerade as learning. Any world built from familiar concepts lets prior knowledge do the work.
- 01FrameFind the valuable decision, workflow or experience.
Our best early result was wrong. We said so.
The first runs showed the protocol layer cutting agent communication by about three quarters. It would have made a great headline. We took it apart instead and found the cause: the protocol condition used shorter message templates. We normalised them, the advantage mostly vanished, and we wrote it down. That reframed the programme: the hard problem isn't building the layer. It's proving it.
- Retracted a 74% communication cut as an artifact
- Normalised templates across every condition
- Listed six flaws in our own earlier design
- 02DesignShape the interaction, context and control model.
A world no model has seen before.
The next phase, GLYPHWORLD-R, generates universes with randomised physics and nonsense vocabulary, so no model can lean on what it already knows. Some rules are never observable at all: only an agent that has formed a real abstraction can get them right. A cosmetic decoy catches superstitions, and three roles each see part of the physics, so the full picture only exists if they build it together.
- Randomised physics in every universe
- Held-out rules that only true abstraction can solve
- Three roles with split sight: knowledge must be pooled
- 03EngineerBuild the intelligence, agent and tool chain.
Thirteen ways to be wrong, tested side by side.
Instead of one comparison, the design runs a full set of arms that isolate each part of the idea: no memory, transcript search, search over the layer's own history, append-only notes, a curated shared layer that agents can add to, support, contradict, revise and merge, private layers per role, a placebo, a perfect rulebook, a one-shot summary, a reader from a different model family, alternating writers, and lesions that remove parts of the layer to see what breaks.
- Curated layer ops: add, support, contradict, revise, merge
- A reader from a different model family tests vendor independence
- Placebo and lesion arms to catch false credit
- 04ProveEvaluate performance, safety and economics.
Decide what counts before you look.
The statistics are pre-registered: 20 to 28 universes, paired tests with bootstrap confidence, correction for multiple comparisons, equivalence tests so a null can be a real answer, and an early futility check. The analysis code is validated on synthetic data before any real run. The design itself won a unanimous vote of a three-judge review council against four competing proposals.
- Pre-registered analysis with a decision for every outcome
- Analysis validated on synthetic data first
- Chosen unanimously over four competing designs
- 05OperateDeploy, observe and improve in the real workflow.
Frontier questions on a startup budget.
The definitive run is costed at roughly $210 to $280 of model calls, with a hard cap of $350 and about four weeks of work. Fast, cheap, falsifiable: the kind of research loop that lets a small team ask a frontier question and get a real answer.
- Hard budget cap of $350 for the full study
- About four weeks from build to answer
- Every outcome, including null, already has a next step
The answer · An experiment that can't lie
The first thing a research programme should get right is how it will know it's wrong.
- 13Experimental arms
- 6Flaws in our own work, designed out
- 1Headline result we retracted
- $350Hard cap for the definitive run
05What this could unlock
What this could unlock.
- 01If it holdsAgent fleets with institutional memorySupport, operations and engineering agents that inherit what earlier agents learned, so the thousandth task is handled better than the first.
- 02If it holdsKnowledge that outlives the modelKnow-how stored outside any model's weights, readable by agents from different vendors, so switching models doesn't mean starting over.
- 03If it holdsAuditable machine knowledgeEvery pattern and strategy carries its evidence and lineage, so teams can see why an agent believes something, and retire it when it stops being true.
- 04If it holdsCoordination across boundariesAgents that each see part of a supply chain, a network or an organisation, pooling structure so the whole system acts on a picture none of them holds alone.
- 05If it holdsLearning without retrainingImprovement that happens in a shared layer in minutes, instead of a fine-tuning cycle measured in weeks and budgets.
- 06If it holdsHumans and agents, one commonsA layer that people can read, correct and contribute to, turning agents' experience into knowledge the whole organisation shares.
06Glimpse of the research
A glimpse of the research.
07FAQ
Questions about Asymmetri.
Yes. Asymmetri runs an internal research practice alongside client work. Project Protocol is its research programme into shared machine cognition: whether teams of AI agents can store what they learn in a shared, inspectable layer outside any model's weights that later agents inherit and improve. Asymmetri built a deterministic simulation and experiment harness to test it and designed a pre-registered next phase with thirteen experimental arms.
In Asymmetri's research, a protocol layer is an external store of structured machine knowledge: named situation patterns, strategies with evidence of when they helped or hurt, failure corrections and the lineage of how each idea was revised. Agents write to it, curate it and inherit it, so knowledge can compound across agents and sessions instead of being lost when a session ends or a model changes.
Retrieval-augmented generation and vector databases store text or past transcripts and fetch similar passages. Asymmetri's protocol layer stores structure: typed patterns and strategies with usage evidence, a lifecycle and lineage, retrieved by an auditable structural vocabulary. Asymmetri tests it directly against transcript search at the same reading budget, because a shared layer only matters if it beats plain retrieval.
Asymmetri compares against strong baselines and ceilings, not just against doing nothing, and decides what counts before looking. In Project Protocol that meant equal-budget transcript search, an all-seeing oracle ceiling, placebo and lesion arms, worlds with randomised rules so models can't rely on prior knowledge, and pre-registered statistics. When an early 74% communication saving turned out to be an artifact, Asymmetri retracted it.
Yes. Asymmetri builds production multi-agent systems and researches how they can accumulate knowledge. The same disciplines from Project Protocol (structured shared memory, auditable lineage, fair evaluation and cost caps) shape how Asymmetri designs agent platforms for clients, from research agents to delivery tools where AI agents act on a person's behalf.
Asymmetri is a hard problems technology company. It engineers AI systems, agents and digital products for global brands, across four areas: intelligence, process, experience and scale. Asymmetri works through four practices: AI agentic systems and data, application development, technology consulting, and innovation and R&D.
Every Asymmetri engagement runs one loop with five stages. Frame: find the valuable decision, workflow or experience. Design: shape the interaction, context and control model. Engineer: build the intelligence, agent and tool chain. Prove: evaluate performance, safety and economics. Operate: deploy, observe and improve in the real workflow. The loop repeats, so each pass lands closer to the target. Asymmetri's case studies show each hard problem moving through it.
Send a short brief through asymmetri.in/contact-us or email hey@asymmetri.in: the problem, what's in the way and your timeline. No deck or RFP is needed. Asymmetri replies to set up a framing conversation, then proposes a prototype or pilot with scope, team, timeline and what success looks like.
08More work