Engram
The physical trace a memory leaves
Each memory is a distilled claim that records which agent wrote it, when, and from what source. Ask what you believed in March and get a real answer.
Memory for agentic AI
One memory for every AI you use. It stays with you.
MAEVE sits beneath your agents, models and machines, so what one learns, all of them know. Your memories live on hardware you choose, outside any single provider.
Any agent · Any model · Every machine
Coming soon
In private beta and in daily use today.
Ownership
Memory is too important to belong to one app.
Every AI tool builds its own picture of you, and none of them share it. Change tools and what the old one learned stays behind. MAEVE moves that knowledge out of the tool and into your hands.
Your data, your disk
Memories sit in databases on hardware you control. No cloud account is required and nothing phones home. A full backup is one command.
Any agent, any model
MAEVE speaks MCP, the open standard for AI tools. Claude, Cursor, OpenClaw and anything else that supports it share one memory. Retire a tool and nothing is lost.
Every machine
Each machine keeps a local replica, so reads are instant and work offline. Changes sync through a small server you run yourself.
How it works
Designed the way memory works.
Brains link, carry, insulate, consolidate and forget. MAEVE does each of these for your AI.
Synapse
Where two neurons connect
New memories link to related ones, and people, projects and devices form a knowledge graph. A question can reach the right memory when no keyword matches.
Axon
The long-range signal
Changes travel from one machine to the rest through your own sync server. What your laptop learns, your desktop knows.
Myelin
Insulation that keeps signals apart
Personas keep your worlds separate. A client session never opens your family or home memories, and a broken setting stops recall instead of widening it.
Consolidation
What sleep does for memory
Each night MAEVE files conversations by topic, distills durable facts, merges near-duplicates and flags contradictions for you to settle.
Pruning
Letting unused connections go
Memories nobody uses lose confidence over time and retire, which keeps recall sharp. What you pin stays, and nothing is silently destroyed.
Recall
Recall you can measure.
MAEVE applies state-of-the-art retrieval methods and tests them against LongMemEval, a public benchmark for long-term memory. An exact lane finds file names and settings first. Full-text, term, vector and graph signals then fuse into one ranking, packed to fit the model's context.
- 94.9%
- of questions had the right conversation turn in the top 10 results
- 91.3%
- had it in the top 5, with query expansion on
- 5
- retrieval signals working together
LongMemEval-S, held-out split, with real embeddings. These figures measure whether the turn holding the answer was retrieved. They do not measure end-to-end answer accuracy.
Learning
Experience carries over. Private details stay put.
Lessons
A fix learned for one client can help the next. MAEVE rewrites it without names or details, and nothing is shared until you approve it.
Experience
Agents log why they acted and what happened, then check that record before acting again. Past failures surface first.
Your approval
A move that would change who can see a memory waits for you. Anything else MAEVE files on its own is logged and can be undone.
MAEVE
Coming soon.
Your tools will keep changing. What your AI knows about you should stay, and it should stay yours.