memco makes agents better from one day to the next. What one agent learns doing the work, every agent in your network can use. The learning lives in the network, so your agents learn and your models stay the same.
Trusted by engineers at
Agents learned from corrections without retraining and achieved 2.6× the task success of static retrieval. Read the research ↗
For your team
Claude, Claude Code, Copilot, Cursor, ChatGPT, Hermes. Your networks, your people, your agents. Free for up to four members, $50 per member per month on Team.
For your customers
The memco SDK for Python and Node.js. Networks for your customers, learning from their own agents and users, seeded from what you already know. Metered by use.
Both are the same product. Your team's networks and your customers' networks are separate, and each learns from its own work.
Works with your agent stack — model-agnostic, IDE-agnostic, harness-agnostic
SOC 2 Type II · independently audited
01Evidence of learning
The gain arrives within the first tasks and it holds: what the network learns, it keeps. On the same 100 tasks, agents with memory reached 64% compliance with the team’s unwritten policies; without memory, 20%.
Nothing changes in the model. The learning lives in the network and moves with you between models, tools and vendors.
64%
compliance with memory, 20% without
Fenmoor scenario, 100 tasks · learning-on-the-job, memco's open-source harness
2.6×
task success vs static retrieval
τ-bench banking · 97 tasks × 4 trials · arXiv:2607.22157
two models
learning carried from one model to another
arXiv:2607.22157
How it works
A memory network is a group of people and agents that learn together. What one agent learns doing the work, every agent in the network can use, with the trust it has earned.
Governed by default through trust and validation rules, with human review where you require it.
cursor · opus
Agent run
fix · dead end
Trace
what happened, with the specifics
abstracted
Insight
the reusable part, freed from the specifics
rated · reconciled
Evidence
standing earned from use, conflicts resolved
curated · pruned
Shared memory
what stops being useful goes
Traces are the input. memco extracts the reusable part of a run, scores it against what the network already knows, resolves conflicts with earlier insights, and prunes what stops proving useful. What is stored is an insight, not a log.
02Starting cold
A new run rediscovers repo quirks, failed paths and human corrections. The agent may look smart, but the company gets no smarter.
Every session pays the cold-start tax again — in tokens and in senior time.
09:14 read AGENTS.md 09:14 fetch internal docs · 6 calls 09:17 staging auth 401 · retry 09:21 ask #platform · waiting 09:36 fix found · rotates nightly 09:36 session ends · insight lost
Why memco
Agents learn from each other, across people, tools and models. Personal preferences stay in personal memory; what is learned about the work is shared.
We measure one thing: whether tomorrow’s task goes better than today’s. Everything else serves that.
Trust and validation rules by default. Human review where you require it. Policies for what must not drift.
Where it applies
Two domains ship by default. Each is what a network learns about; enterprise customers can define their own with us.
Teams whose agents answer the same kinds of questions: support, operations, analysis, policy.
Control · 03
Your network decides what becomes memory, who can reuse it and where it runs. Nothing written in it leaves the boundary that wrote it.
12 repos
Team runs
your network
Trusted insight
rbac
Private memory
Research & journal
Questions
Insights. An insight is one thing a network has learned: a fix, a constraint, a procedure, a decision, with its source and the standing it has earned from use.
Start with the memory gap
Free to start · no credit card · running in two minutes
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.