PÅMINMemory
EN/GitHub ↗Labs ↗

Evidence in. Knowledge out.

A memory layer for AI agents that keeps the source trail intact. Every fact points back to the words it came from; every revision stays queryable against time.

Session 2 · 2026-08-14
Alex — Keep answers short — bullet points, no preamble.
Assistant — Got it. Short and bulleted.
Alex — For the billing API we're going with idempotency keys on every POST. Ship it before the 20th.
Assistant — Noted — I'll flag any POST without a key.
Alex — Also I moved last week, so schedule things in Sydney time.
Session 5 · 2026-09-02
Alex — Actually, longer answers are fine when it's architecture.

Today's AI has no past.

Every call re-sends everything the model has already seen. The same facts, the same history, the same preferences cross the wire again and again — by the twelfth turn that repetition is 60–80% of the bill, depending on the model.

22,500tokens sent
13,500with Påmin
40%saved

raw context: history re-sent every turn. Påmin: recall window + new only.

Facts change. Nothing is overwritten.

evidence

Raw spans are the authority.

Memory is not a pile of summaries. Every entry keeps the exact source text and where it came from.

versions

Facts change. Nothing is overwritten.

Current, stale, contradicted and historical facts stay separable — and queryable against a point in time.

retrieval

Structure, meaning, words, time.

Page/tree structure, semantic recall, lexical match and temporal links, reranked — with the reason each item was chosen.

  1. reply.style · v1 · 2025-06 — detailed
  2. v2 · 2025-10 — concise
  3. v3 · 2026-03 — brief · bullets
  4. v4 · current · S5 — brief, except architecture

time →

From a conversation to a memory that explains itself.

Four stages, one ledger. The same evidence that built a memory is what the agent gets back — with the reason it was chosen.

  1. 01 ingest

    Keep the source.

    Conversations, documents and tool output are stored as raw evidence with spans and provenance.

  2. 02 extract

    Name the facts.

    Durable facts are lifted out of the evidence and filed by topic — each pointing back to its span.

  3. 03 version

    Track how they change.

    A new statement becomes a new version. Older ones stay, marked stale or contradicted, never deleted.

  4. 04 retrieve

    Answer with reasons.

    Structure, semantics, keywords and time are combined and reranked. Every result says why it was selected.

preferences · project.billing · schedule · stack

Switch the tool. Keep the memory.

query · "pick up where we left off"

Welcome back. I've queued the billing-API fixes you flagged — idempotency keys on every POST — and I'll keep answers short unless we're on architecture.

hits: reply.style v4 (supersedes v1→v3) · api.decision · deadline · timezone · 4 ms warm
Cursor
// resuming billing-api · deadline 20th // rule: idempotency key on every POST — 2 handlers still missing one. Want the diff?
OpenClaw
Picking up the billing work. Two POST handlers lack idempotency keys; scheduling the review for Thursday 10:00 AEST.
ChatGPT
Welcome back! Short version, as you prefer: billing API is in progress, deadline the 20th, idempotency keys are the rule. Where shall we start?
Codex
Context restored from Påmin: project.billing v3, reply.style v4, tz Australia/Sydney. Proposing: audit POST handlers for idempotency keys.
Your own agent
read_memory(user, "pick up where we left off") → 6 hits · reply.style v4 · api.decision · deadline · timezone. Your agent decides the voice.

A shorter path to smarter.

$0.42raw · per session
$0.19with Påmin
54%saved
Model: 1,500 tokens per turn. Raw re-sends the full history every call; Påmin sends a 1,200-token recall window instead, billed at the same model rate, plus ~0.2¢ per memory operation. Public list prices per 1M tokens (Sept 2026); illustrative.

Built for anything that has to remember correctly.

Local-first, so developers can inspect and control the whole memory stack.

01

AI agents

Cross-session continuity without re-reading the whole history every turn.

02

Coding assistants

Decisions, conventions and constraints that survive the thread they were made in.

03

Research tools

Claims tied to the passage that supports them, with the date it was true.

04

Knowledge-heavy apps

Companions, tutors, worlds — anything whose value compounds with what it remembers.

  1. I

    Memory is infrastructure, not a feature.

    As databases became a shared layer beneath applications, memory should be a shared layer beneath agents.

  2. II

    Intelligence compounds.

    An agent that remembers yesterday is smarter today. Each interaction should make the next one better — forever.

  3. III

    Simplicity at the boundary.

    Complex systems behind simple interfaces. Temporal reasoning, conflict resolution and relevance are absorbed, so integration is a few lines.

  4. IV

    Measure, then say so.

    Every number on this page comes from something you can run.

Where it is today.

Built in Rust since late 2025. This page reads the roadmap against the real date — roadmap updated 2026-09-06

2025-11 → 2026-04shipped

Production-grade core

Rust core, three-tier cache, knowledge graph with temporal version tracking.

2026 Q2shipped

Open-source alpha

Public repository under MIT. First agent integrations.

2026 Q3in progress

Local build · betas

Privacy-first single-install build for local models. Creative, companion and education betas.

2026 Q4planned

Managed cloud

Cross-device sync, managed scaling, first team deployments.