MimesisA memory you own
Every AI you
use forgets you.
Or worse — it remembers wrong, and quietly carries the mistake into everything it does next.
Mimesis gives you one memory you actually own — your facts, your preferences, the way each of your assistants should behave. You decide what goes in. Every AI you connect gets exactly that, and nothing you didn't choose.
Bring an Obsidian vault, a folder of chats, or nothing at all. Your data stays yours: export or delete it, whole, at any time.
Nothing becomes memory until you say so.
The repetition tax
You re-introduce yourself to a machine that was talking to you an hour ago.
Every new chat, every new tool, every new agent — you re-explain who you are, what you're working on, how you like to be spoken to. The context you built up yesterday is gone.
And the memory features that do exist quietly decide things for you: an assistant infers a “fact,” saves it, and repeats it back for months. You never approved it. You can't see where it came from. You can't take it back.
Mimesis flips both of those. One memory, portable across every AI. And you — not the model — decide what's true.
How it works
Four steps between “who are you again?” and an AI that already knows.
Bring what you have
Connect an Obsidian vault, import past chats, or just answer a short set of questions about yourself. Originals are preserved, untouched.
Keep what's true
Mimesis and your assistants propose memories. You approve, edit, or discard each one. Suggestions never become facts on their own.
Get the right context
For any task, Mimesis compiles the handful of things that matter — not a wall of history — sized to fit whatever model you're using.
Share it on your terms
Connect Claude, your own agents, anything that speaks MCP. Each one sees what you allow, in the voice you set for it.
Works with your agents
Built for the multi-agent tools you already run.
Mimesis speaks MCP, the open standard these systems are adopting, so your one memory shows up right inside them — no copy-paste, no starting over.
See the full list and how to connect.
You hold the pen
A fact isn't a fact until you keep it.
When an assistant thinks it's learned something about you, it doesn't just save it. It waits in a review queue. Keep it, reword it, or throw it out — and every kept memory can be traced back to exactly what it came from.
- Nothing an agent infers becomes durable without your yes.
- Edit a memory in your own words before it sticks.
- Delete anything, and it stays deleted and auditable — never silently resurrected.
Your agents, your team
You build the team. Mimesis is its shared memory.
A developer, a marketer, an executive assistant. More and more real work happens across a team of focused agents. You create those agents where they run. Mimesis is where each one's memory, persona, and settings live, all drawing on the same picture of you. Write a persona, bring a SOUL.md, or let a short interview draw it out, then keep sharpening it as the agent works.
- You create the agents; Mimesis holds and helps you refine each one's memory, persona, and permissions.
- They share one memory of you, so nobody starts from scratch and nobody drifts apart.
- Scoped and safe: a teammate carries only the persona you gave it, and sub-agents propose to your memory without ever overwriting it.
The right context, not all of it
It hands the model what matters — and shows its work.
Dumping your whole history into a prompt is expensive and makes answers worse. Mimesis compiles a small, relevant briefing for each task and tells you what it included, what it left out, and why. It runs locally and deterministically — no guessing, no data sent off to be “embedded” somewhere you can't see.
- Prompt-aware: the memories that fit this task, not a transcript.
- Fits any model's context window — you set the budget.
- Every briefing is explainable: included, excluded, and the reason.
$ mimesis context "prep me for the supply-chain review" --budget 1200 Briefing (5 kept · 2 trimmed · fits 1200 tokens) · you prefer terse, decision-first summaries · product = Monte-Carlo reorder-point engine · review is with the ops team, not engineering Left out full chat archive · stale profile notes — not relevant to this task
Yours, and only yours
A memory this personal should belong to you.
Your memory lives in your own private store. Retrieval runs locally — nothing about you is shipped off to an external model or embedding service to work. Obvious secrets are refused before they're ever saved. And you can walk away with everything, or erase it, whenever you want.
How we handle your data →Isolated by person
Each owner's memory is a separate store. No shared pool, no cross-tenant bleed.
Export & delete, whole
Take everything with you as clean files, or delete it for good. It's your call, always.
Provenance kept
Originals are preserved so any memory can be traced back to where it came from.
Secrets refused
API keys and passwords are blocked at the door — they never land in your memory.
Who it's for
If more than one AI knows a little about you, they should agree.
People who live in AI
You use a handful of assistants every day and you're tired of repeating yourself. One memory, and they all start from the same page.
Builders running agents
You run your own agents and want each to have a stable personality and a memory you curate — not one that drifts as it invents facts.
Teams that need a paper trail
Shared context is useful right up until it's a liability. Keep memory owned, reviewed, and auditable across the people who rely on it.
Straight about where we are
Mimesis is early and we'd rather show you the real thing than a stock-photo dream. The engine works today across a command line, an HTTP API, and MCP, with imports, review, personas, and local retrieval. The hosted beta is where you come in.
Start one memory today. Bring your AIs to it.
Free to start. No credit card. Export or delete everything whenever you decide it's time.