We often hear the claim that if you put enough powerful AI models together, collective intelligence will “emerge” and produce a 1+1>2 effect. In practice, it’s often the opposite: 1+1<2—sometimes even chaos, confusion, and internal friction.
The old story of the Tower of Babel captures this perfectly. People tried to unite to build a tower to the heavens, and God derailed the project with a single move: he scrambled their languages so they could no longer communicate. Today, AI teamwork is running into a new kind of “Babel” crisis.
Memory is the scarce resource
Why do multi-agent collaborations fail? At the root is memory scarcity.
In the single-agent era, we focused on “context engineering.” Think of it as one person’s working memory: useful, but limited in capacity. Whether a model has a 128k or 1M context window, it’s still tiny compared with the ocean of information it must operate in.
In multi-agent systems, this scarcity is multiplied. You can’t cram every agent’s conversations, every historical decision, and every intermediate variable into every agent’s head. It’s not just expensive—tokens cost money—it’s also that attention is scarce. Too much noise causes overload, and agents stop knowing whom to listen to.
This is exactly what many AI teams experience:
The PM agent updates requirements, but the dev agent keeps coding against the old spec.
The QA agent reports a failure, but can’t tell whether it’s a new feature or an old bug.
The whole team falls into “context jitter.”
Fundamentally, this is an issue of information transaction costs being too high. If every agent must synchronize through massive volumes of dialogue, the cost of collaboration becomes prohibitive.
Memory engineering: lowering the transaction costs of collaboration
Economics teaches that when transaction costs are too high, you need institutional design to reduce them. In AI collaboration, that institutional design is “memory engineering.”
You can think of memory engineering as the AI team’s external neocortex. Instead of relying on each agent’s meager RAM (temporary memory), it builds a shared, persistent hybrid of ROM (durable, canonical memory) plus RAM (working state).
Without this unified system, multi-agent setups exhibit classic “distributed split-brain” failure modes:
Duplicate work: what one agent has already researched gets researched again because the information wasn’t shared—pure waste.
State inconsistency: agent A thinks the project is paused while agent B continues to push full-speed ahead because updates weren’t synchronized—coordination failure.
Cascading hallucinations: one agent’s mistake becomes another agent’s “truth,” and the error amplifies through the chain—systemic risk.
To fix this, you need a mechanism that reduces transaction costs by establishing clear “ownership” and identity for memory.
Give memory an “ID card”
In human society, we invented property deeds, contracts, and invoices to reduce transaction costs. In AI collaboration, we need to turn unstructured conversation into structured “memory blocks.” Every piece of information written into the shared brain should come with an “ID card.”
A memory block can’t be just text; it must include metadata—similar to how market transactions need precise coordinates, owners, and timestamps, not vague descriptions like “something over there.”
A standard memory block should include:
Timestamp: to prevent stale information from contaminating decisions—like an expiration date; you don’t drink expired milk.
Source: clearly who said it. A PM’s statement about requirements should carry more weight than a developer’s casual guess—clear accountability and authority.
Relations: what other information is it linked to? Build a knowledge graph so agents understand logical dependencies between facts.
With this “ID card” system, collaboration stops being headless, ad-hoc thrashing and becomes rule-based, efficient exchange.
Three pillars of a shared brain
At the engineering level, building this “shared brain” requires three pillars. Each maps to a different function of the human brain—and to different foundational infrastructure in an economy.
1) Semantic retrieval engine
This is memory’s “hippocampus,” and also a high-efficiency search engine. Agents don’t need exact keywords (e.g., “JWT auth failure”). They can describe the issue fuzzily (e.g., “Why did we abandon that login approach?”) and still retrieve the relevant memory. This dramatically lowers information search costs.
2) Knowledge graph
This is memory’s “associative cortex,” and the logistics network of information. When an agent queries “login,” the system can follow the dependency chain and warn: “Note: login depends on the user service, and the user service changed its API yesterday.” This connectivity prevents “blind men and the elephant” local optimization.
3) Immutable event log
This is memory’s “time chain,” and a strict accounting ledger. Whoever changed the code or altered a decision must record it immutably. This isn’t only for accountability—it gives agents “time travel” capability: roll back to an earlier decision point and re-simulate reasoning. It protects historical assets. This is a classic application scenario for blockchain technology.
Closing
The evolution of software development is, in essence, the history of continuously reducing the cost of collaboration. From single-machine programs to large distributed systems, from single agents to agent swarms, the core challenge never changes:
How do you deliver the right information, at the lowest cost, to the right entity, at the right time?
In the future of AI development, architecture will become central again. We won’t just design prompts; we’ll design the “universal language” and the “shared storage structure” between agents.
If prompts are an agent’s “mouth,” then memory engineering is the agent team’s “brain.” Only with a solid memory foundation will the AI team’s Tower of Babel avoid collapsing under the noise of collaboration.
Questions to think about
In human teams, what mechanisms function as “memory engineering”? If you remove these mechanisms (e.g., banning documentation and meeting notes), how would collaboration efficiency change?
Finally, if you don’t have a computer background but want to learn about the latest progress in AI, this book is recommended: AI from Scratch (“Zero-Basics AI”), https://github.com/yeasy/ai_beginner_guide. It uses engaging stories and case studies to help you get started quickly.
No comments:
Post a Comment