Use casesAn individual

One developer, three clients, nothing re-explained.

Tomas works alone from Porto for three clients, in Claude Code, Cursor and Codex, on three codebases with three sets of rules. One personal network followed him between tools and models. Three client networks never touched. The first reuse came on day two and the free allowance ran out in week three.

People: One contract developer
Clients: A payments start-up, a logistics firm, a publishing house, each with its own repos and tracker
Agents in use: Claude Code for long tasks, Cursor for daily edits, Codex where a client's budget sits with OpenAI
Tools connected: GitHub, Linear
Memory networks: Four: one personal, one per client
Plan: Free for the first fortnight, then Team at $50 a month
Time to first proven value: Day two

01Starting point

Where Tomas started

Tomas bills three clients a week: a payments start-up that rejects any pull request without a changelog line, a logistics firm running a Postgres fork with its own migration tool and a deploy freeze published in the on-call rota, and a publishing house where nothing goes live without editorial review. He uses Claude Code for the long tasks, Cursor for the daily edits and Codex when a client pays for OpenAI. Each client has its own repos, its own Linear workspace and its own unwritten rules.

Every session started from zero. The agent did not know about the changelog line or the migration tool. Tomas re-typed the same context into each tool, and again into each new model when one shipped. He kept a notes file per client and pasted from it a dozen times a day. A correction he made in Cursor on Monday was unknown to Claude Code on Tuesday, and a rule the logistics client had drilled into him was unknown to Codex on Wednesday.

He counted a week in March before he installed anything. Roughly a fifth of billable hours went on re-explaining what the agents had already been told, and two pull requests that week were rejected for mistakes he had corrected the week before. The agents were not getting worse. They were simply never getting better at his clients, and he was the only memory in the system.

02Day one

Day one

An afternoon, most of it spent deciding what belonged to which client. No client changed anything on its side and none of them knew the plugin was there until he told them.

Claude Code, Cursor and Codexmemco plugin in each, one account, one login. Ten seconds per tool.
GitHub and LinearPull request outcomes and tickets attach to insights as evidence, in the network of the client whose repo they belong to.
SeedThe three notes files became thirty-one insights, each scoped to a client or to Tomas himself. Lines he could no longer vouch for were left out.
SparkFour networks drawn in spark.memco.ai. The personal network is the only one every tool reads in every session.

PersonalHow Tomas works: branch naming, commit size, the standard library for dates, how he wants pull requests described.Every agent Tomas runs, in every tool, for every client.
Client: the payments start-upThe changelog rule, the services on the audit freeze list, the review habits of its lead engineer.Tomas's agents when working in that client's repos.
Client: the logistics firmThe migration tool and its dry-run flag, the deploy freeze window, the on-call rota.Tomas's agents in that client's repos only.
Client: the publishing houseThe editorial review gate, the content model, what never publishes directly.Tomas's agents in that client's repos only.

fig. 01Scope map

The personal network holds how Tomas works and is read in every session. Client networks are sealed from each other. Nothing promotes: a lesson that names a client lives with the client.

03The first weeks

The first weeks

Week one was mostly seed and re-explaining anyway, because the habit of pasting took longer to break than the plugin took to install. The first lesson learned from live work came on day one and was reused on day two in a different tool.

fig. 02Deployment timeline

Spacing is proportional to elapsed days. The filled mark is the first reuse: a correction made in Cursor applied by Claude Code the next day.

Day 1. The payments client rejects a pull request for a missing changelog line, in Cursor.Every pull request in the payments repos includes a changelog entry under the current release heading before it is opened.Day two: Claude Code, on a different ticket for the same client, adds the entry unprompted. First reuse, across tools. The rejection has not recurred.
Day 3. An agent writes a plain SQL migration for the logistics firm and the dry run fails.Logistics migrations are written for the firm's own migration tool, never as plain SQL, and are run with the dry-run flag first.Scoped to the logistics network. The next three migrations run clean on the first dry run.
Day 6. Codex, working for the publishing house, applies the payments client's changelog rule to a content repo that has no changelog.A rule learned in one client's repos applies there only. The changelog rule is scoped to the payments client.The insight had been saved to the personal network by mistake. Tomas moves it to the client network. He adds one rule for himself: a lesson that names a client lives with the client.
Day 9. An agent proposes a date library in a payments service.Use the standard library for dates in every project. Do not add a date dependency.Saved to the personal network, because it is Tomas's rule, not a client's. Read by every tool for every client from then on.
Day 12. A deploy to the logistics firm lands inside the freeze window and the on-call engineer asks why.Logistics deploys happen outside the freeze window in the on-call rota. Read the rota before any deploy command.The rota is linked as provenance. No deploy has landed in a freeze since, in any tool.
Week 5. The payments start-up moves its agent budget from Claude to OpenAI.The client network is read by Codex exactly as it was by Claude Code.Nothing is re-explained. The lessons were in memco, not in a model's context, and the model switch cost an afternoon of setup instead of a fortnight of re-teaching.

04Governance

Governance

A lesson that names a client lives with the clientClient networks are read only when Tomas is working in that client's repos. The personal network holds how he works, never what a client does.
Nothing crosses between clientsNo client's network reads another's, and nothing from any client is promoted into the personal network. The wrong-scope mistake on day six is the reason the rule is written down.
One identity, one audit trailEvery read and write is logged against Tomas's account, so a client who asks what the agents knew about their work can be shown it.
Retirement when a contract endsWhen an engagement closes, its network is archived and its memory exported for the client in a portable format if they want it.

fig. 03The write path

The lesson is scoped to the payments client's network. Every tool Tomas runs in that client's repos reads it; no other client's session can.

05What compounded

What compounded

Tomas kept the same count he had made in March and made it again in May and July.

Metric	Movement	Why
First-pass acceptance	55% to 81%	Pull requests accepted without a change requested, across all three clients by month two. The payments client's lead engineer asked what had changed.
Repeat-error rate	Two a week to none	Repeated rejections, counted over eight weeks. The last repeat was the day-six scope mistake.
Time to an accepted result	A fifth to a twentieth	Time spent re-explaining, as a share of billable hours. The pasting habit took a fortnight to die. By week three the notes files were closed.
Cross-person or agent learning gain	Every lesson, another tool	A correction made in Cursor was applied by Claude Code or Codex within days, every time.
All-in cost per accepted result	Tokens per PR down a third	Counted after reads, which added about five percent to each session. Fewer retries and shorter corrections removed more.
Improvement with experience	Unchanged on a model switch	When a client moved its budget to OpenAI, the memory followed, because it was never in the model.

fig. 04Two counts

Pull requests accepted first time across three clients, and billable hours spent re-explaining, as counted in March and again in May.

fig. 05Costs before it pays

Share of billable hours spent re-explaining context, against the week counted in March. The pasting habit took a fortnight to die.

06What it cost

What it cost

Setup timeOne afternoon. Three plugins, one login, four networks, thirty-one seeded insights from the notes files.
Who ran itTomas. About ten minutes a week in Spark in month one, mostly moving lessons to the right network.
Plan and tokensFree for the first fortnight. The 1,000 operations a month ran out in week three with three clients and three tools, so Tomas moved to the Team plan at $50 a month. Against roughly $350 a month in agent spend, the token saving alone covered it.
What it replacedThree notes files, a dozen pasted context blocks a day, and the re-explaining that followed every model switch.