MIRA remembers things across your conversations today — your name, your preferences, the projects on your mind. But I’ll be honest: its memory is good, not great. Making a personal assistant’s memory genuinely reliable — so it recalls the right fact, doesn’t quietly forget, and doesn’t confidently make something up — is a hard, open problem.
It’s a problem I’ve been researching, and I’ve just published three preprints on it. They don’t describe features MIRA ships today — they set out why memory is hard and what a better foundation looks like, and they’re the direction MIRA’s memory and model work is heading.
The three papers
1. Addressing Is Not Free — Factored Knowledge Storage and the Key-Discrimination Limit of Dense Models. When you pack facts into a model’s weights, retrieving them isn’t free: dense models hit an “addressing wall.” Careful accounting shows they store far less than commonly assumed — roughly 0.3 bits per parameter, not the ~2 often quoted — a limit that simply scaling the model can’t overcome. The paper proposes a factored alternative: a small reasoning kernel plus a separate, addressable store. (DOI: 10.5281/zenodo.22059102)
2. Surfaces Do Not Reach Addresses — Across ~94,700 real-world entities, addressing by an entity’s canonical name resolves almost perfectly (0.9996), but asking by an alternative surface form or description reaches the right entity only ~4.75% of the time. The takeaway: reliable recall needs explicit lookup indexing, not retrieval the model is merely hoped to have learned. (DOI: 10.5281/zenodo.22059900)
3. A Dial, Not a Line — On certifying that a store returns the right answer: it introduces “refusal with a certified price” — an answer-or-refuse gate that trades coverage for a measured error rate, so a system can abstain instead of hallucinating when it isn’t sure. (DOI: 10.5281/zenodo.22060127)
What this means for MIRA
To be clear about where things stand: these are research foundations and a future direction, not features MIRA has today. MIRA’s current memory works and will keep improving in the ordinary way in the meantime.
But this is the work aimed at taking MIRA’s memory from works to great — toward an explicit, addressable, auditable memory store that recalls the right thing and can honestly say “I’m not sure” rather than inventing an answer. That’s the kind of memory a private assistant your family relies on deserves.
All three preprints are open access on Zenodo — read, cite, or critique them. More to come as the work continues.