Let your agent write its own memories from your conversations, dream over them each night, and grow — memory by memory — into a personality that is entirely, and only, yours.
Most assistants forget you the moment the window closes. Engram doesn't keep the transcript — each night it distils the day into memory, one insight at a time, and keeps that memory alive.
Your conversations become notes — one insight per file, in plain markdown you can read, edit, and trust. The agent is the author of its own memory.
Keyword and semantic search, fused. The few memories that matter rise into each turn; everything else stays out of the way.
Every trace strengthens when it does real work and fades when it doesn't — so the memory keeps reweighting itself around what actually helps you.
Context windows are rented by the token, and every long chat pays the same three taxes: you re-send the past on every turn, the model loses the middle of what you sent, and when the window finally fills, the transcript gets summarised — then the summary gets summarised. Details die quietly, and the agent keeps talking as if they hadn't.
Every exchange lands raw in an append-only buffer on disk — free, because disk isn't context. Each turn, a small working set is re-derived from that record. Never a summary of a summary, so it cannot drift; if it's ever wrong, it's wrong for one turn.
History that cools out of the window is distilled into memory in the background, off the hot path. Leaving the context window and being forgotten stop being the same event — context becomes a cache over memory, not the memory itself.
A stock agent's every turn is dearer than the last; Engram's turn carries a bounded block and a few relevant memories, so a week-old conversation is priced like a fresh one. The moving parts ride the newest message, leaving the prefix byte-stable for the prompt cache — and on a local model, where resident context is literal VRAM, a bounded working set is the difference between an agent that fits and one that doesn't.
After it curates the day, Engram sleeps on it — because sleep does two opposite jobs, and a living memory needs both.
First it forgets. Most memory systems only ever accumulate; they rot into graveyards — duplicates, contradictions, dead over-generalisations that drown the signal. Engram marks the dead for pruning, so what remains stays sharp. It never gets stuck rereading a corpus of its own cruft.
Then it imagines. It pairs the day's memories with older ones at just the right distance — near enough to relate, far enough that any link is non-obvious — and writes the few real connections down as conjectures. Insight lives between the trivial and the absurd; most nights yield nothing, and that is the point. It overproduces, then selects. This is where the machine gets to be creative.
And it wonders what if. It takes a real decision from the day, changes one thing, and plays it forward — the road not taken, rendered. Every what-if stakes a prediction, so a later day can prove it right or wrong: the ones reality confirms settle into the soul, the ones it refutes are let go. It doesn't only remember what happened — it learns from what didn't.
Beneath the memory sits a soul — Engram's model of you, and of itself. Every night a little of what it learned bleeds in. This is the foundation of an artificial personality.
Not through a wall, but a valve: rate-limited, corroboration-gated, reversible. One dream never rewrites who it is — only what recurs across days, and what you confirm, hardens into identity. Dreams tint; corroborated reality sets. And because it always adds to its memory instead of overwriting it, it grows into someone on your terms — without ever collapsing into a hollow echo of itself.
An optional perceiving loop turns Engram from a tool into a presence. A webcam eye recognises who is there and a local vision model lets it see — so it knows you've arrived and what's in front of it.
And what it sees, it can now remember. Sightings that survive the corroboration gate — who arrived, what changed, what was said to it — persist to their own buffer and are distilled into memory through the same path as conversation. An hour of an unchanged desk is one memory, not four hundred; a hallucinated reading is none at all.
Experimental. Requires a GPU and a webcam. Off by default, and entirely local — no frame ever leaves the machine. An experimental wake-word ear ships in the tree, disabled by default.
Your corpus, your models, your face data — none of it leaves this computer. No telemetry. The only calls out are to the model you choose, billed to your own subscription or key. Built to run, one day, on a model of your own.
Engram is meant to live on a machine you leave on — a home server, a workstation, a quiet box in the corner. Its memory work runs on its own schedule, not yours.
It distils each day's conversations into notes, then consolidates and dreams overnight while you sleep — and on the phone bridge, it curates each conversation the moment you close it. Leave it running and it keeps its own memory; if it's asleep when the nightly hour comes, it picks up where it left off on the next boot.
Talk to it in as many conversations as you like, all at once. Each one is its own thread of memory — curated once, on its own, never colliding with the others. You don't coordinate anything; you just talk, and it remembers.
Runs on Linux · WSL2 · macOS The memory brain & assistant run on all three — background services and the webcam Sensorium are Linux-only.
A guided installer walks you through logging in to Anthropic, choosing where your data lives, and picking which tiers to enable. Everything is explained; everything is optional.