Memco for Knowledge Work
Knowledge work is where judgement is the asset.
“Vendor blog posts are leads, not evidence — two independent sources before a claim ships.”
The task ships, the correction lands — and the judgement disappears into the chat.
It becomes a sourced, scoped lesson the next agent reuses — across models and tools.
One agent learns. The next starts ahead.
Private by default · Scoped to the right work · Sourced to the original evidence · Owned by you
See one correction on Monday improve a different agent’s work on Friday.
Live demo · the second task · Monday → Friday
Different task. Different agent. Same hard-won lesson.
Synthetic demo data · runs in this page · no model call
Different task. Different agent. Same hard-won lesson.
Memco sits between your agents as the learning layer: one private, domain-aware memory, connected once over MCP.
Recognisable work
Four workflows, the same pattern: what happened, what Memco learned, what the next agent does differently.
Evidence
In a controlled banking evaluation, learning from corrections reached 2.6× the single-trial success of a static-RAG baseline. The memory transferred between Mistral Large and Claude Sonnet 5 in that setup.
Memco controlled evaluation· not a customer deployment resultRead Learning on the JobActivation has one definition here: the first later task that visibly benefits from a prior lesson. You measure it on your own work, not ours.
The example calls your own private Memco space, so it needs a free signup first.
What Memco keeps
Chat history and retrieval preserve what was said and what is written down. Neither turns a correction into something a different agent can act on next week.
The expensive part is not storing context. It is paying to rediscover the same judgment.
How a lesson becomes trusted
A correction is not a lesson yet. It becomes one when it has a source, a scope and enough evidence to be validated — and it stops being one when the world changes.
correction or outcome
Learn
Something you fixed, decided or proved becomes a candidate lesson.
source · scope · evidence
Govern
Memco attaches where it came from, where it applies and what makes it trusted.
a different task, later
Reuse
A future agent applies it; feedback updates or retires it.
Source, scope, status and retirement are visible objects on every lesson — not fine print. Without them, memory does not compound. It becomes sludge.
Ownership
Everything your agents learn lands in a space you own — private for you, shared when you invite your team.
Every lesson carries its source: whose correction, which task, what evidence made it trusted.
Lessons apply where they belong — a client, a workflow, a team — not everywhere at once. Private by default; nothing is published anywhere outside your space.
A conflicting lesson surfaces for review instead of silently merging. Stale knowledge loses trust and is retired — struck through, not quietly deleted.
Disconnect a client, revoke access, take your lessons with you. Change model, agent or interface without throwing away what use has taught your agents.
OAuth 2.0 · API keys for automation · encrypted in transit and at rest · regional and data-residency options subject to plan
One connection, separate domains
One MCP connection. Separate Coding and Knowledge Work stores. A repository fix never leaks into a client-reporting workflow.
Connect once
Memco is the learning layer between them, not another place to do the work. One secure MCP endpoint, no plugin per tool, nothing re-explained.
Plans
Prove the learning loop on your own work first. Move to organisation-owned controls when it becomes production-critical.
A private space for one person proving the learning loop on their own work.
Free at launchsubject to fair-use terms
Private space · both domains · MCP access
Organisation-owned learning across people, tools and workflows, with shared review. Free for up to 4 people.
Org scopes · shared review · free to 4 people, then paid
Governed, deployed learning for larger or regulated agent programmes.
VPC and on-prem deployment options
Full seat pricing and trial terms on Pricing
Questions
Start with the next repeated lesson
Connect one agent, correct one task and see whether the next related task starts ahead.
Free to start · No public sharing required · MCP setup in minutes
Need organisation controls? Scope a pathfinder →
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.
