Påmin Memory
Evidence in. Knowledge out.
A memory layer for AI agents that keeps the source trail intact — versioned facts, explainable retrieval, local-first. Rust.
memory.paminlabs.comThe parts agents are missing.
Påmin Labs is an applied research studio. We build infrastructure for AI that has to work over time — a memory that persists, a team that assembles — and publish the measurements with the code.
What we believe.
- I
Agents need infrastructure, not features.
Memory, delegation, review, boundaries — the parts a person takes for granted at work are missing beneath today's agents. We build them as shared layers, not as add-ons.
- II
One agent is not a team.
A second opinion before you commit, a reviewer who did not write the thing, someone holding thirty open threads — capability comes from structure, not from a bigger model.
- III
Safety is isolation, not instruction.
An agent that cannot reach your machine, your real tokens or your remote does not need to be told to behave. Boundaries live in containers and code — prompts get forgotten by turn twenty.
- IV
Measure, then say so.
Every claim on these pages points at a number the repository can reproduce. Where a thing has not been done end to end, we say that too.
Two instruments.
Both open source under MIT. Both built to be inspected.
Orchestrator
An AI team for a company of one.
Say it, approve the plan, merge. Ten roles build, test and review inside containers you never open. One binary.
orchestrator.paminlabs.comResearch directions.
The questions behind the two products — and what a third one would have to answer.
- Q1
How should an agent hold a fact that used to be true?
Representing facts that change: versions, contradictions, and what was true when.
- Q2
Can retrieval say why, from the same evidence it drew on?
Returning not only the right context but the reason it was chosen — from the same evidence ledger.
- Q3
Where does trust live when ten agents share one repository?
Multi-agent systems where trust, ownership and budgets are enforced by the runtime, not requested from the model.
- Q4
What does done mean when the author is a model?
Deterministic gates and reconciliation so that "done" is decided by exit codes, not by the agent's summary.
Three founders.
Based in Sydney. Small on purpose.
- Jason[email protected]Founder · systems, architecture & product
Product, architecture, design and the low-level implementation. Optimisation and systems by background.
- Co-founder · product tuning
How the tools feel in someone's hands, and what to leave out.
- Arvin[email protected]Co-founder · machine learning
Retrieval, ranking and the models behind memory.
Open source, in the open.
Both repositories are public on GitHub. Roadmaps on each product site read against the real date.
Write to us.
Research partnerships, early integrations, or just a question.