Project Protocol
What if machine knowledge could compound outside the model?
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
- Practices
- AI Agentic Systems and Data, Technology Consulting, Innovation and R&D
- Year
- 2026
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.
03Insight
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.
04Hard 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
- Search 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.
- Remembering 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.
- Agents 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.
Through the loop
Frame
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.
Design
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.
Engineer
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.
Prove
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.
Operate
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.
Back to Frame
If knowledge can compound outside the weights, every fleet of agents gets an institutional memory.
- Headroom: oracle vs today's agents
- 1.9×
- Logged episodes
- ~150
- Conditions, incl. an oracle ceiling
- 4
- Token read budget per agent
- 180
08Hard 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
- Proxies flatterEasy metrics like message length or token count can move for reasons that have nothing to do with the idea being tested.
- Builders leak knowledgeIf the people building the system hand it the answers, even indirectly, it will look like it learned them.
- Models already know thingsA language model's training can masquerade as learning. Any world built from familiar concepts lets prior knowledge do the work.
Through the loop
Frame
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.
Design
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.
Engineer
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.
Prove
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.
Operate
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.
Back to Frame
The first thing a research programme should get right is how it will know it's wrong.
- Experimental arms
- 13
- Flaws in our own work, designed out
- 6
- Headline result we retracted
- 1
- Hard cap for the definitive run
- $350
What this could unlock.
Project Protocol is early research, and its answer isn't in yet. But the question is one of the largest open questions in applied AI, and if the hypothesis holds, here is what changes.
- Agent fleets with institutional memorySupport, operations and engineering agents that inherit what earlier agents learned, so the thousandth task is handled better than the first.
- Knowledge 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.
- Auditable 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.
- Coordination 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.
- Learning without retrainingImprovement that happens in a shared layer in minutes, instead of a fine-tuning cycle measured in weeks and budgets.
- Humans and agents, one commonsA layer that people can read, correct and contribute to, turning agents' experience into knowledge the whole organisation shares.
All case studies
AI platform · Web + mobile
Prescribe Life
An AI performance-intelligence platform for leadership coaches. Behavioural science, biometric data and an AI co-pilot flag burnout risk between sessions.
Read Prescribe Life
AI platform · Commercial planning
Komerz OS
An AI commercial-planning platform for consumer brands. A research agent answers where to play, then a planning workspace builds how to win, down to a P&L that reconciles.
Read Komerz OS
Health platform · WhatsApp + web + doctor app
Quicli
Doctor consultations on WhatsApp. Say 'Hi', describe your symptoms in your own language, and a real doctor's prescription arrives in the same chat, with medicines and tests ready to order.
Read Quicli
Internal platform · Web app + AI agents
FinishLine
Asymmetri's own delivery operating system: projects, QA, clients, chat and calendars in one secure app, where AI agents work as the person who sent them.
Read FinishLine