Second Brain — Persistent Memory for an AI Collaborator
An AI collaborator that accumulates a corpus from its own work — and stops being a prompt.
- •Two-layer memory: an explicit cloud corpus plus an implicit persona layer
- •2,346 memories extracted, typed and scoped out of real work — no curated dataset, no hand-maintained prompt
- •What repeats stops being a memory and becomes a principle: 68 active, reusable across projects
- •The working personality is grown, not written — 58 of its 64 traits came from the sessions themselves
- •Same base model, measurably different behaviour — the corpus is the difference
Tap to enlargeEvery AI assistant starts each conversation as a stranger. You re-explain the project, re-state the constraints, re-paste the context — and the moment the session ends, all of it is gone. I built the opposite: a persistent memory system where an AI collaborator accumulates knowledge from its own working sessions and carries it forward. Nothing is stored as a transcript. Raw working context is captured, classified and scoped into typed memories; what recurs across projects is then distilled into a reusable principle; and the handful of memories about how to work — style, boundaries, what not to repeat — accumulate into a working personality that is loaded before the first message of a session is read. After 81 sessions that is 2,346 memories across eight projects, 68 principles standing above them, and a persona layer where 58 of its 64 traits were grown by the work rather than written by me.
Two Layers, Like Human Memory
Nobody authored the personality — 58 of its 64 traits precipitated out of the work
The system separates memory the way people do. The explicit layer is a cloud corpus — typed, scoped facts that are retrieved on demand and matched to the request at hand, so only what the moment needs is surfaced. The implicit layer is different: a small set of memories about identity, working style, and boundaries that is injected unconditionally at session start, before the first message is even read.
What makes that layer interesting is where it came from. I wrote six of its traits by hand. The other fifty-eight the system extracted from the work itself — how I want disagreement handled, which corrections are permanent, where the line between honesty and agreeableness sits. Nobody authored the personality; it precipitated out of the sessions.
That distinction is the whole design. One layer is recalled. The other the assistant simply already is by the time the conversation begins.
The Corpus Writes Itself
Captured from real work; what repeats stops being a memory and becomes a principle
Nothing here is a curated dataset or a prompt maintained by hand. Hooks capture facts out of the actual working day — decisions, constraints, lessons, preferences — classify them by type and scope, and store them automatically. A reflection pass then looks for what repeats across projects and promotes it into a reusable principle: 68 active, 56 more waiting on evidence.
The loop closes at the next session start, when the relevant slice comes back. Every session leaves the corpus larger than it found it.
Observability for a Mind
An invisible memory store made operable — gaps and drift visible at a glance
Memory you cannot inspect is memory you cannot trust. Every session opens with a status card — how much is accumulated and where, which principles are active, whether local and cloud are in sync, and what landed in the last few days.
It turns an invisible, slowly-drifting store into something operable: gaps are visible, sync failures surface immediately, and the growth of the corpus is a number rather than a feeling.
Why This Stops Being a Prompt
The model is off-the-shelf — the accumulated corpus is the part that can’t be copied
A prompt is written by a person and stays where it was put. This corpus is generated by the work itself, changes behaviour without anyone editing it, and transfers across projects — a constraint learned during a security review surfaces months later, unprompted, in an unrelated build.
The practical result is the point: the same underlying model behaves measurably differently depending on the corpus behind it. The intelligence is off-the-shelf; the accumulated context is not, and it is the part that cannot be copied over a weekend.
Interested in building something similar?
Let's Connect