Knowledge Base Server (Memstalker)
A shared memory server that lets Claude, Codex and Gemini read and write the same knowledge base.
Systems
How the systems connect
Six steps, one loop. Each step below says what it does, how far along it is, and which projects make it up. Select a step in the picture to jump to it.
- Knowledge baseMemstalker · MCP
- ModelsClaude · GPT · Gemini
- AgentsDaniel · Agent Swarm
- Harness & evalsGates · reviews
- ApplicationsFour Presient demos
- LessonsMistake ledger → KB
1
Knowledge base
Decisions, fixes and lessons are captured once, with their source, and every agent searches the same store before it works.
Memstalker Side-by-Side Demo
The same prompt built twice: once by a stateless AI, once by an AI with a knowledge base plus agent reviews.
2
Models
The model does the reasoning and can be swapped. CareOps runs the same questions on four models to compare cost and accuracy.
CareOps
A healthcare-operations AI assistant where the software, not the model, decides what it can see and do.
Presient Skill Evals
Published results on whether optimized AI skill prompts beat the originals, failures included.
3
Agents
Agents get roles, fail over when one runs out, and work a task graph in waves, with a watcher and test gates between waves.
Daniel (Agent Orchestrator)
A terminal orchestrator that fails over between Claude, Codex and Gemini so work continues when one runs out.
Agent Swarm
Keeps a team of coding agents working a task graph, with reviews and test gates between waves.
Source not public yet Case study
4
Harness & evals
Stage gates, independent reviews and tests decide what counts as done. Blind evaluations check whether a change really helps.
Presient SE Harness Pack
A test rig measuring whether a software-engineering harness helps an open-weights coding model.
Source not public yet Case study
Presient Skill Evals
Published results on whether optimized AI skill prompts beat the originals, failures included.
5
Applications
Bounded agents in realistic products, where server code holds the permissions. CareOps and the Browser Agent run live in Docker behind Traefik on a Linux server; the Website Chatbot and Bid Selection run locally.
CareOps
A healthcare-operations AI assistant where the software, not the model, decides what it can see and do.
Browser Agent
Watch a constrained, read-only AI browser agent work a fake records site, with every blocked action shown.
Website Chatbot
A Shopify store chatbot that answers only from the store's own content and says when it doesn't know.
Bid Selection Memo
Turns a synthetic construction bid package into a cited GO / HOLD / NO-GO memo a reviewer must approve.
6
Lessons
In the agent swarm, review findings are recorded. Recurring ones, once approved, become warnings on the next matching task.
Then back to step 1: the lesson is stored in the knowledge base, where the next task finds it.
Agent Swarm
Keeps a team of coding agents working a task graph, with reviews and test gates between waves.
Source not public yet Case study
Knowledge Base Server (Memstalker)
A shared memory server that lets Claude, Codex and Gemini read and write the same knowledge base.
Where it runs
Real infrastructure, run by me
The live demos run in Docker containers behind Traefik on a Linux server, with HTTPS certificates from Let's Encrypt and DNS on Cloudflare. Source lives on GitHub, and every change goes through tests before it is deployed.
Designed, not running yet
Software Factory
The whole loop in one visible workflow: a ticket for a fictional company goes through requirements, knowledge-base lookups, a plan, the build, tests, an independent review and a human approval. The private workflow has run one practice job end to end, the Website Chatbot. The public version and its benchmark are designed but not running yet, and nothing is shown until it has really run.