Case study

Ten months in production: what one team’s memory became.

By the Firekeep teamPublished Updated

Firekeep is the shared memory system one development and operations team has run every working day for more than ten months. It was not built as a demo. It was built because the same problems kept getting solved twice.

The snapshot

November 10, 2026. One team, one deployment.

267active operational skills, up from 45 in August
33internal documents ingested, up from 12
2.4memory reads per session, averaged over the 50 most recently evaluated sessions
845 / 64commits and release tags as of August 9, and still shipping

The incident that proved it

Mid-deployment, an AI assistant went silent, with no useful error to go on.

The cause was the kind that costs an afternoon: Python was resolving through a symlinked release directory while the pipeline copied only part of the required configuration. Once found, the fix was captured as a skill.

Weeks later, a different session touched an adjacent area. The skill was already in its pre-flight briefing.

Nobody searched for it or remembered it, and the problem never got a second afternoon.

A library nobody sat down to write

The skills cover deployment recovery, Kubernetes access, certificates, Jira and agent-tool configuration, DNS and LDAP, Kafka, and CI/CD. Every one was captured at the moment a problem was solved, by the agent that solved it. That is why the library kept growing while the team was busy doing other work.

One memory, two teams, no handover

Development and operations share a single memory. Operations inherited development’s design decisions without a handover meeting. Incidents worked in operations became skills that development saw later, without anyone forwarding a thing.

Documents that find you

Ingested runbooks and guides are split into searchable sections. The relevant part surfaces during recall, so nobody has to remember which file it lives in, or that the file exists at all.

Entirely on the team’s own infrastructure

Firekeep runs on a Rancher-managed Kubernetes cluster with Redis, Neo4j, Qdrant and local Ollama inference. There are no per-token API charges, and no data leaves the building.

What ten months taught us

Install Firekeep See the recall benchmark Why a file is not memory