memco journal
Memory, measured.
Research, engineering practice and field notes on how agents learn from real work — and how that learning becomes reusable company memory.
Latest writing.
- When companies run on agents, how do they work together?Mergers and partnerships have always relied on people making sense of each other's businesses. As agents take on more of the work, that understanding needs somewhere to grow.
- Who's teaching our agents?At memco, our agents learn from three sources: ourselves, our colleagues and our team's agents. Which do you think makes the largest contribution?
- Shared Memory, now for your customersThe memco SDK for Python, Node.js and Go puts Shared Memory inside your product: agents that learn how each of your customers works, in memory networks you design, with knowledge your customers can inspect and override.
- The moat of the agentic companyHow startups, scaleups and enterprises can build a defensible advantage from organisational learning as capable AI agents become widely available.
- Learning on the Job, Batteries IncludedAn agent learns a business's unwritten rules and compliance climbs from 20% to 64%. The harness behind our Learning on the Job paper is now open source, so you can reproduce that curve for about $3.
- Introducing Knowledge WorkKnowledge Work is a new domain in Spark for the corrections, decisions, procedures and client rules that business work produces all day. Lessons keep their source and scope, survive a change of model or tool, and are tested by whether the second task gets better.
- One private memory for every agent you useBuilder is now free for one person and Teams is free for up to four. One private Spark memory, outside any model or application, that follows you across the agents you use and starts earning its place on the second task.
- Memory for Knowledge WorkYour teams correct AI assistants every day, but that knowledge disappears with the session. Spark's new Knowledge Work domain turns those corrections into governed organisational memory every assistant can reuse.
- Sovereign AI Needs Sovereign MemorySovereign AI is more than controlling where the model runs. New research shows why organisations should also own the governed, portable memory that gives agents their operational knowledge.
- Learning on the Job: Agents Can Learn from the Feedback They Already GetEvery business has feedback its AI agents can learn from. New research shows that external memory turns routine corrections into portable, compounding operational knowledge.
- What should one AI agent be allowed to teach another?Shared memory lets a second agent skip a dead end. It can also pass on a mistake. What should cross between agents is the smallest reviewed unit of judgment, with evidence, scope and a write path that decides what the organisation has earned the right to trust.
- Coding agents are the wedge. Organizational memory is the prize.Coding agents are the cleanest place to see the agent-memory problem: every run creates evidence, but without trusted organizational memory the next agent starts cold.
- Knowledge Management Systems Are Always Obsolete. Agents Can Fix ThatEvery enterprise knows its knowledge base is out of date. AI agents can finally fix that, but only if the memory infrastructure meets enterprise requirements. We walk through what it takes: identity and access control, knowledge scoping, human oversight, provenance, and data residency.
- Why Active Agentic Memory is the Next ShiftHuman-crafted knowledge works perfectly fine, until data-driven learning surpasses it. Explore the historical "Zig and Zag" of AI, and why I believe shared agentic memory is the infrastructure required for the next era of autonomous learning.
- Your Team Knows More Than Anyone On ItMost AI memory tools give back what you put in. With Knowledge Abstraction, Spark derives principles your team never stated, and helps agents avoid problems nobody has encountered yet.
- When the model changes, does the company keep learning?A day of degraded models and blown usage caps showed engineering teams how dependent they are on one provider. Fallback models answer where the next token comes from. The harder question is whether the company keeps learning when the model changes.
- Claude memory is good. It is not the same thing as organisational memory.Anthropic's memory for Claude Managed Agents validates the category and is not the same thing as organisational memory. Platform memory is local state; companies need a neutral layer where what the organisation learned survives across tools, models and time.
- Your Agent's Memory Is a Markdown File. That's a Problem.Files are the most popular form of agent memory. Here are four things they structurally cannot do.
- Your Coding Agent Remembers Everything, Until It Doesn'tContext windows can retrieve facts with remarkable accuracy, until compaction silently erases the project rules and conventions your team depends on.
- Continual Learning for Enterprise AI Needs a Memory LayerMost enterprise AI systems do not fail because the model is incapable. They fail because the system cannot retain, refine, and reuse what the organization has already learned.
- World Models Need MemoryJack Dorsey and Rohit Krishnan both describe a future run on company world models. Neither says where the operational knowledge comes from. The memory layer underneath, built from validated corrections and shared across teams and vendors, is the load-bearing wall.
- Making Sense of Agentic Memory: A Map of the Design SpaceThe term 'memory' is used to describe wildly different systems. We map out the design space and explain the choices behind Spark's shared memory architecture.
- If Your Coding Agents Don't Share Memory, You're Burning MoneyMost teams running AI coding agents pay for the same knowledge over and over. We ran 200+ evaluation runs to show how shared memory cuts token usage up to 87% and lifts pass rates from 70% to 100%.
- Your Agents Don't Need Better Models. They Need a Shared Brain.Every enterprise will run agents in every function, and each one will know nothing about what the others learned. The bottleneck is not model intelligence but organisational memory: knowledge that travels between teams, in privacy tiers, captured by observation rather than documentation.
- What Happens When Your AI Agent Fails?What happens when an agent fails? We analyzed the ROI of Spark shared memory and discovered that costs decrease by 40% on average. The pleasant surprise was the 'fail cheap': even unsolved tasks see a cost reduction of 34%.
- Agents need knowledge they can't generate themselvesA new paper confirms what we've been building toward: curated procedural knowledge improves agent performance by 16%, but agents can't write it themselves. Static skill files don't scale. Shared memory does.
- AI-Generated Code Might Be More Maintainable Than You ThinkA preprint on AI and software maintainability finds no significant difference in downstream maintainability—and habitual AI users may produce slightly cleaner code. Caveats and takeaways.
- Why Shared Memory Matters Even for "Solved" ProblemsWhen an agent already solves a task 100% of the time, does memory help? Our experiment with bug-fixing in Go shows Spark cuts cost and variance in half—even on the easy problems.
- Why AI Struggles with Your Legacy CodeThe gap between AI demos and real engineering teams: legacy code, internal APIs, and tech debt. How a shared agentic memory closes that gap with collective continual learning.
- Beyond the Human CeilingAI is surpassing human performance in specialized domains. In software development, agentic knowledge-sharing infrastructure like Spark can outperform the human forums it replaces.
- Smarter Together: Our New Paper Shows How Shared Memory Lifts All ModelsOur first evaluation paper demonstrates that access to Spark's shared, active memory benefits AI coding agents of all sizes—from smaller open-weights models to the largest commercial models.
- The Collective Intelligence GapWe're building an agentic workforce. Now we need to build the infrastructure for collective intelligence.
- The Wedge and the Vision: Code to Agentic InfrastructureLLMs broke software's most powerful feedback loop — learning from distributed experience. Shared memory can fix it. Our path from code agents to the runtime for agentic systems.
- 79% of Enterprises Are Deploying AI Agents. Only One Thing Is Missing.Enterprises succeeding with AI aren't the ones with the most powerful models. They're the ones that figured out how to make organizational knowledge compound through private shared memory networks.
- What Developers Actually Want From AI Memory: Lessons from 230+ HN CommentsThe Hacker News response to Claude's Memory launch reveals a sharp divide: sophisticated users want sovereignty, control, and learning - not accumulation. Here's what the community is demanding.
- Karpathy is Talking About Cognitive Core, We're Building the Memory for ItAndrej Karpathy's vision of separating intelligence from knowledge aligns with our R&D at Memco. Our Spark memory layer enables smaller models to achieve state-of-the-art performance by externalizing specialized knowledge.
- Memory Incentives and ArchitecturesAs AI performance gains slow, memory becomes the new battleground. We explore service-provider, portable, and community memory architectures, and how they incentivize learning and ownership.
- Nature's Blueprint: What Human Memory Teaches Us About Building Smarter AI AgentsBy studying biological memory, we can design more effective memory systems for AI agents that learn, generalize, and strategically forget
- Reinforcement Learning For AI Agents: Learning on the Job with Active MemoryHow an active memory system enables runtime reinforcement learning, transforming AI agents from static tools into dynamic partners that learn from experience
- The Art of Forgetting: Why True Memory is the Next Frontier for Autonomous AI AgentsBuilding truly autonomous agents requires moving beyond simple retrieval to design active memory systems that can learn, generalise, and strategically forget
- Introducing Memco: the shared memory layer for AI agentsEvery day, thousands of developers and their AI agents solve the same problems—then forget the solutions. We're building Memco to fix that.
- Agents That LearnWhy collective learning and shared memory are the missing pieces for AI agents
- Your Next User Is Not HumanThe Developer Flywheel is Broken. How AI agents are replacing humans as the primary consumers of developer tools and documentation.
Nothing in this section yet.
Go deeper
The research behind the writing.
the loop · benchmarks · product · research
A short dispatch on shared memory for AI agents.
The numbers behind the product, what we’re shipping, and the research we’re reading. No filler.