Velqua · local-first memory experiment

The problem

Most AI sessions begin with no useful context from the work that came before. Existing memory products often solve that by asking each application to integrate a new SDK or depend on a hosted service. Velqua explored a smaller, local-first boundary.

What I learned

Storage is not the difficult part. Relevance, correction, provenance, and forgetting are. A memory system that retrieves the wrong fact at the wrong time can be worse than no memory at all.

The project therefore treats memory as a governed context layer rather than a transcript archive. Useful context should be explainable, bounded, and removable.

Where it fits

continuity

Carry selected context between local sessions without copying entire transcripts.

relevance

Prefer context connected to the current task instead of injecting everything available.

control

Keep data local and give the operator an explicit way to inspect or remove it.

interoperability

Keep the application-facing boundary small enough for different local tools to use.

The public boundary

The public project demonstrates the product idea and the engineering approach. I do not publish private memory contents, ranking heuristics, agent identity rules, internal coordination flows, or deployment details.

Velqua was an important step toward a broader principle I still use: persistent context needs governance, not just more storage.

Inspect the public repository on GitHub. Read the companion reflection: what shared memory taught me about agent coordination →