The story
Long before the systems and the papers, there was a reason I care this much about memory.
Read the full story →Charles Weeks · AI memory systems
Building AI systems that remember.
Twenty-five years in residential systems integration, finding the hard-to-find solutions. Lately spent building AI memory systems that run local-first and hold up.
The Spire Library keeps the work that counts: the systems, the papers, the build logs.
Selected work
Built, deployed, in use.
01
Command center / current
The Observatory
A private command center built over a plain markdown vault. Publishing pipeline, session memory, and an exemplar loop that learns my voice from my own edits. It runs at home on my own hardware, and it is where Alita, the assistant, lives.
More about The Observatory →02
Memory substrate / current
Personal Brain
Local-first memory for an AI that actually knows its operator. Retrieval tuned for recall that holds up months later. Nothing leaves the building.
More about Personal Brain →03
Anchor paper / published
Sovereign Knowledge Federation
The SKF paper: durable AI memory with capture, structure, decay, and recall treated as one system. Published as a v1.0 working paper, with a DOI.
Read the paper →Shipped · open source
The code is public.
spire-skills
Two plugins: spire-pipeline, where a design swarm, a run-list, and a conductor keep a multi-session AI build from losing the plot, and model-vetting, which reads a HuggingFace model repo before you pull it, quick triage first and a deeper static pass when it matters. I run the same shipped versions myself.
sovereign-knowledge-federation
Source and issues for the SKF paper, kept as a living repo instead of a frozen PDF, so corrections and questions have somewhere to land.
The foundation
Before the AI work, twenty-five years of high-end residential technology. Networks, control systems, and the operational discipline of making complicated systems behave inside real homes for clients who do not accept excuses. Along the way I hired and trained dozens of technicians, ran the field, and helped build the systems that keep the operation running. That is the reason the AI work ships instead of staying a demo. I have done the unglamorous part before.
Recent writing
The latest from the archive.
Sovereign Knowledge Federation
The anchor paper: durable AI memory with capture, structure, decay, and recall as one system. Published as a v1.0 working paper.
Every directory needs a janitor
Append-only context rots. The reconciliation pass is the discipline that keeps the flywheel from grinding itself apart.
Nobody is building systems
Harness engineering in three layers: the Brain thinks, the Muscle executes, the Playbook governs. The payoff is coherence over time.
A better harness, not a better prompt
Two weeks of prompt tweaks could not fix what one architecture change did. Splitting one overloaded pass into four focused ones.
Connect
The Library is open.
If you are building something that has to remember, need someone who finds the answers that are hard to find, or just want to compare notes, I am glad to talk.
Start a conversation