Påmin Memory — questions and answers

Answers about the memory layer: what it stores, what it returns, and what it costs. The roadmap on the home page reads against the real date.

1. What is Påmin Memory?

   Påmin Memory is a plug-in memory layer for AI agents that keeps the source trail intact. Every fact it stores points back to the exact words it came from, and every revision of that fact stays queryable against time. It is written in Rust and local-first, under the MIT licence at github.com/pamin-labs/PaminMemory. It is a v0 alpha: the repository carries the README and the design notes today, not yet the engine.

2. How is this different from a vector database or plain RAG?

   A vector store returns passages that are similar to a query. Påmin Memory returns facts with their evidence and the reason each one was selected. Retrieval combines four signals — page and tree structure, semantic recall, lexical match, and temporal links — then reranks them. The result carries a citation back to the source span, so an answer can be checked rather than trusted.

3. What does "facts change, nothing is overwritten" mean?

   When a new statement contradicts an older one, the older one is not deleted. It is marked stale or contradicted and kept. Current, stale, contradicted and historical facts stay separable, and any of them can be queried against a point in time — so you can ask what the system believed last March, not only what it believes now. This temporal version tracking is the part most memory systems do not have.

4. Which agents and models does it work with?

   Any of them. Påmin Memory is a layer beneath the agent, not a feature of one model. The same memory serves Claude Desktop, Cursor, OpenClaw, ChatGPT, Codex or an agent you wrote yourself — switch the tool and the memory follows. Your agent decides the voice; Påmin decides what it gets to remember.

5. Does my data leave my machine?

   Not in the local-first build. It is a single install that keeps the whole memory stack — evidence, facts, index — on the device, so developers can inspect and control it. A managed cloud tier with cross-device sync is planned separately; the roadmap on the home page says where each tier stands.

6. How much does it actually save on tokens?

   The saving comes from not re-sending the whole conversation every turn. Raw context grows with the square of the conversation: turn twenty re-sends nineteen turns of history. Påmin sends a fixed recall window plus what is new. On the model shown on the home page — 1,500 tokens per turn, a 1,200-token recall window, plus roughly 0.2 cents per memory operation — a twelve-turn session sends about 72% fewer input tokens. What that is worth in money depends on what the model charges: roughly 59% on Claude Opus 4.1, 54% on Claude Sonnet 4.5, and close to nothing on the cheapest models, where the per-operation fee cancels out the token saving. The calculator on the home page recomputes both for the model and length you pick; the numbers are public list prices, not a benchmark.

7. What language is it written in, and what is the licence?

   Rust, with a three-tier cache and a knowledge graph that carries temporal version tracking. MIT licence, public repository.

8. Where is the project today?

   The production-grade core and the open-source alpha have shipped. A privacy-first local build for local models, with creative, companion and education betas, and then a managed cloud tier with cross-device sync, follow. The roadmap section on the home page reads status.json against the real date, so it is accurate on the day you load it rather than the day it was written.