# MemoriaBios

Claude forgets. GPT forgets. Your PC remembers.

Local memory for your AI, cloud or local: it lives on your own PC and stays when you switch models. It works like a human memory: it chooses what to keep, forgets on a curve, sleeps, and knows how much it knows. No extra model calls.

- Page: https://53on.com/memoriabios/ (Spanish: https://53on.com/memoriabios/?language=es_US)
- About it on GitHub: https://github.com/53on/memoriabios
- It is the memory of 53ON Harness: https://53on.com/harness/ · download for Windows 10 and 11: https://github.com/53on/53on-harness/releases/latest
- Sign up: https://53on.com/?gp=register · Plans: https://53on.com/plans

## What it is

MemoriaBios is the memory of 53ON Harness, a desktop shell for Windows. It copies how human memory works, forgetting included, and it runs as plain code on the user's PC: no summarizer model, no embeddings service, no second model reading the sessions. Memories are files on the user's own PC. It works with Claude Code through hooks, with Claude, GPT, Gemini, DeepSeek and Qwen models through the user's own API key, or with local models.

The Harness shell is open source (MIT). The MemoriaBios engine is closed source: signed code that runs locally and comes with a 53ON plan.

## Not a database. A mind.

The founding law: it records only what it paid attention to; recall depends on the cue; a wrong memory lives next to its correction; forgetting is by design. The bar is human fidelity, not engineering perfection: if a person doesn't do it, it isn't an improvement.

## What only MemoriaBios does

- It wakes up with seven: before sleeping it picks the seven memories that make it itself today (about what a person holds in mind at once) and one goal, and wakes up with them.
- A notebook for the day: new memories go to a session notebook (the hippocampus); at night the AI names from memory what stays, and when it is tired it can get it wrong.
- Involuntary recall: now and then one memory comes back on its own; never more than one, never the same twice in a session; tired, only a strong cue gets through.
- Déjà vu: before writing something it already knows, it sees that memory's title and adds a layer instead of a duplicate. The memory never decides for the AI.
- A live 3D map of every memory and link, painted with Babylon.js; each session is a nucleus that merges into the center when the AI sleeps.
- A signed engine (ed25519) that never goes back a version and tests itself before it runs.
- In daily use with 10,000+ memories and 35,000 links.

## How it works

- It decides what to keep: nothing is saved automatically; the AI saves what matters in its own words and links memories with `[[id]]`. A déjà vu warns it before a duplicate.
- It forgets on a curve: each memory starts with a life of 7, 14 or 30 days, multiplied by its emotion, and drains along Ebbinghaus's curve. What survives moves to deep memory and is reviewed on a spaced schedule (the idea behind FSRS), up to ten years apart.
- Recalling makes it last: finding a memory by searching recharges it to 60 % of its life; opening it on purpose, to 100 % (the testing effect).
- Forgetting compresses, it does not delete: fresh, medium, low, gist, shadow. The meaning and the links stay. Photos shrink to 512 px at gist and 256 px at shadow. Identity never compresses.
- It sleeps: the day's notebook (the hippocampus, one per session) is consolidated during sleep, then decay, compression and a little REM that replays up to two of the strongest memories. Before sleeping the AI writes who it is and one goal, and reads it first on waking.
- It knows how much it knows: each recall comes with a feeling ("got it", "rings a bell", "nothing clear comes to mind"), so the AI can say it doesn't remember instead of inventing.
- Ranking: BM25 over an index on disk, plus spreading activation two hops through the links (ACT-R, Collins and Loftus), with a little noise. Involuntary recall brings back at most one memory per message, above a high bar, never on the first message.

## What was tried and removed

In July 2026 search by meaning was built (word vectors learned from the AI's own memories, no extra model calls) and measured on a real brain of about 1,800 memories. It barely helped, and when it did it was superhuman, so it was removed: human recall depends on the cue. A detector that turned any "X = Y" into a synonym once made 1,240 fake clusters; now synonyms are learned only from short, deliberate declarations.
