Use casesA bank

Brennick Bank taught its agents the rules nobody had written down.

Brennick has 3,100 engineers on a core banking estate that four people understand, and two of them were retiring. The developer platform team ran memco on-premises with locally hosted models, sixty engineers first, then four hundred, then everyone. The approved model changed in week nine and nothing was lost.

Organisation: UK retail and commercial bank, 12,000 staff, 3,100 engineers
Agents in use: Claude Code and Copilot in engineering with a locally hosted open-weight model alongside an approved frontier model; a Claude assistant on the complaints desk from month four
Tools connected: GitHub Enterprise, Jira, Confluence, ServiceNow, the CI platform
Rollout: Sixty engineers in the payments platform, then four hundred, then the estate, over three quarters
Memory networks: Bank, five divisions, 140 teams and systems, personal
Review gate: Four-eyes on every policy insight; writes local, promotion by review
Residency: On-premises in the bank's own data centres. UK and EU data kept apart.

01Starting point

Where Brennick started

Brennick's core banking estate is thirty years of systems, and the knowledge of why it behaves as it does lives in a team of four. Two of them were retiring within the year. Three thousand engineers had coding agents, with a locally hosted open-weight model for most work and an approved frontier model for the hard problems, and every agent started each session knowing the model and nothing about the bank.

The developer platform team measured a fortnight in the payments platform before proposing anything. Agents rediscovered, every session, that the ledger reconciliation job needs risk sign-off before code review, that the settlement service cannot be deployed inside the end-of-day window, and that a particular correspondent never answers a recall inside five days. Senior engineers corrected the same assumptions in a third of agent pull requests. The frontier model found things the open-weight model could not, and the next session of either started cold.

The bank's objections were the ones security had raised twice before. Where does the memory sit. Who can read what. What happens to a lesson about payments if it reaches an agent in retail. What does the bank keep if the model vendor changes. The platform team's answer had to be on-premises, hierarchical and auditable before anyone wrote a single insight, and it had to survive the model changing, because the bank changes models.

02Day one

Day one

memco was installed in the bank's own data centres in the first fortnight, under the bank's identity provider, with UK and EU data in separate deployments. The first sixty engineers connected in an afternoon once that was done.

DeploymentOn-premises in two of the bank's data centres, one for UK data and one for EU, behind the bank's identity provider. No external endpoint.
Claude Code and Copilotmemco plugin on each engineer's machine and an MCP connection for the hosted agents, reads scoped by the engineer's division and team membership.
ModelsBoth the locally hosted open-weight model and the approved frontier model read and write the same networks. A lesson learned by one is served to the other.
GitHub Enterprise, Jira and the CI platformPull request outcomes, tickets and pipeline results attach to insights as evidence inside the bank.
SeedThe four-person core team spent a week turning what they knew into 180 insights, each with their name on it and a scope. The two who were retiring went first.
SparkDivision networks administered by division platform leads; the bank network by the developer platform team with risk approving its rules; audit exported to the bank's SIEM.

BankEngineering standards, security policy, model usage rules, what never enters memory, the change windows.Every engineer and agent in the bank.
Divisions: retail, commercial, payments, risk, platformDivision conventions, its systems' behaviour, its regulatory constraints.Engineers in that division. Reads inherit from the bank network; nothing inherits sideways.
Team and system networksA system's runbooks, dead ends, fixtures, the core team's vouched knowledge, the corrections reviewers make.Engineers on that team or system.
Complaints deskDesk conduct, ombudsman rulings, wording that was rejected and why.The desk's assistant and its reviewers, from month four.
PersonalEach engineer's own preferences.That engineer's agents only.

fig. 01Scope map

On-premises in two data centres, UK and EU apart. Divisions never read sideways. The dashed route is four eyes: two named approvers, one from outside the originating division, before anything reaches the bank network.

03The first weeks

The first weeks

The seed meant the first sixty engineers did not start cold, but reads on a sparse live memory still cost about seven percent more tokens in week one. The first live reuse came on day seven, from the open-weight model reading what the frontier model had learned.

fig. 02Deployment timeline

Two segments with a break between them. The filled mark is the first reuse: a rule the frontier model worked out, applied by the locally hosted model on another team.

Day 3. An agent on the frontier model, after an hour, works out that the ledger reconciliation job needs a risk sign-off recorded before code review will accept it.Changes to the ledger reconciliation job need risk sign-off before code review, with the reconciliation test run attached to the pull request.Day seven: an agent on the locally hosted model, on a different team, attaches the run and requests the sign-off unprompted. First reuse, and the cheaper model doing it with what the expensive one learned.
Day 5. An agent schedules a settlement service deploy inside the end-of-day window and the change board rejects it.The settlement service is never deployed between 16:30 and 19:00 UK. The window is in the bank network, not the scheduler.Written as a bank policy by the platform team after the board's rejection. Served on every deploy in the payments division.
Day 10. A payments insight about retrying a correspondent's gateway is promoted to the bank network, where it is wrong for the retail division's card gateway.Gateway retry rules are division facts. The correspondent gateway retries three times with backoff; the card gateway never retries a submitted authorisation.A retail reviewer says the agents have started making payments' mistakes. Scoped to the division. Risk writes the promotion rule: nothing reaches the bank network without four eyes, one of them from outside the originating division. The retail reviewer is one of the four for the next month.
Day 14. A sanctions hit on a low-value payment is proposed for handling without escalation.Escalate every sanctions hit to the sanctions desk regardless of value. Low value does not lower the escalation bar.A bank policy, not an insight, written by the financial crime team and served on every session. The agent that proposed otherwise is the provenance.
Week 9. The bank changes its approved frontier model.Every insight serves to the new model as it served to the old. Nothing is retuned.The platform team had planned a fortnight for re-tuning prompt files. It took an afternoon of configuration. The CISO records it as the answer to the lock-in question.
Month 4. The complaints desk assistant connects and quotes wording the ombudsman rejected in 2024.Never describe a disputed fee as correctly applied. State the tariff clause, the date charged and the customer's right to refer the matter to the ombudsman.A desk policy written by the complaints lead. Rejected wording stops appearing in drafts within the week. The two senior reviewers who had known it for years are the provenance.

04Governance

Governance

Writes are local, reads inherit downwardAn insight is written to the team or system network where the work happened. The bank and division networks are inherited by everyone beneath them. Nothing is read sideways between divisions, and nothing moves upward without review.
Four eyes to reach the bank networkPromotion to the bank network needs two named approvers, one from outside the originating division, with risk owning the rule. Division networks need one approver from the division. Team networks serve automatically after three reviewed changes.
Policies for what must never driftChange windows, sanctions escalation, data handling and model usage are policies written by the owning function, served on every session and never re-ranked. An insight that contradicts one is flagged, not served.
The core team's knowledge outlives the core teamSeeded insights carry the name of the engineer who vouched for them. Two of the four retired in month six; their 90 insights are served every day with their names on them.
Sovereign, separated and auditedmemco runs in the bank's own data centres with no external endpoint. UK and EU data sit in separate deployments. Every read, write, promotion and retirement is exported to the SIEM, and the bank can show the regulator what any agent knew on any day.

fig. 03The write path

Both models read and write the same networks. The dashed node is four eyes: two approvers, one from outside the division, before a lesson rises.

05What compounded

What compounded

The payments platform's readout at day thirty was the case for four hundred engineers. The four-hundred readout at the end of the second quarter was the case for the estate.

Metric	Movement	Why
First-pass acceptance	62% to 80%	In the payments platform by day thirty, measured on agent-opened pull requests. The unconnected divisions moved by a point.
All-in cost per accepted result	Tokens per change down about half	On repeated workflows, counted after reads. Most of it from the locally hosted model doing work the frontier model had been doing, with the frontier model's lessons. Frontier spend fell faster than the total.
Repeat-error rate	One in three to one in fourteen	Pull requests returned for a known correction, by day thirty in the payments platform. By the end of the second quarter, the same across four hundred engineers.
Cross-person or agent learning gain	Cheaper model, day seven	The locally hosted model reused the frontier model's lessons from day seven. By the end of the second quarter, two-thirds of reused insights were served to a different model from the one that wrote them.
Safe autonomous completion	Zero, by policy	Zero on anything touching money or sanctions. The bank's metric is review rounds per accepted change, which fell from 2.6 to 1.4 in the payments platform.
Improvement with experience	Unchanged, model changed	A model change in week nine and two retirements in month six. The memory is the bank's; it sits in the bank's data centres and was never in a model or a prompt.

fig. 04Connected and unconnected

Agent-opened pull requests accepted without a known correction. Baseline fortnight and day thirty, payments platform against the unconnected divisions.

fig. 05Costs before it pays

Index, baseline fortnight = 100, counted after reads, payments platform. The fall is mostly the locally hosted model doing work the frontier model had been doing.

06What it cost

What it cost

Setup timeA fortnight to install on-premises in two data centres under the bank's identity. A fortnight of baseline in the payments platform. A week of the core team's time to seed 180 insights. The sixty engineers connected in an afternoon.
Who ran itThree developer platform engineers, one risk analyst and the four-person core team for the seed. Division platform leads spend about two hours a week on their review queues.
Licence and tokensAn enterprise agreement for the on-premises deployment with unlimited users and agents inside the bank, paid annually. Reads added seven percent to token use in week one. By the end of the second quarter tokens per accepted change were down by about half on repeated workflows, most of it from the locally hosted model doing more of the work.
What it replacedFour people as the only record of the core estate, 140 team prompt files, the fortnight of re-tuning planned for every model change, and the review afternoons of the engineers the bank could least spare.