GenArch.
Alexandria · the data-and-knowledge plane

The Architecture of Living Memory

How a durable substrate turns ephemeral context into knowledge that compounds.

This is the canonical deep-dive on how Living Memory is built: the vocabulary, the composition model, and an honest account of what runs today and what is still being built. It is written to be accurate before it is impressive.

Architecture review

Add to the leading wave of AI,
and help steer it for the better.

The work has a primary purpose: to add to the leading wave of AI and help steer it for the better, driven by knowledge, learning, and individual sovereignty. This is the north star. Everything in the architecture below is an attempt to make it real.

Who it serves, in order:

First · Humans
Sovereignty over your own lifelong corpus
Health, education, home, finances, owned by the person who lives them. The individual is the principal; the data is theirs.
Second · Organizations
Compliance, auditability, durable institutional memory
Decision-trace auditability and institutional memory that survives staff turnover. This is legitimate and necessary. Organizations are served in service of the people they serve.
Eventually · Agents
A substrate agents can reason from
That day is coming. The same memory that serves a person can ground the agents that act on their behalf.

This is a mission statement, written to be aspirational. What is actually built is described in the appendix.

Some problems we are addressing.

Three structural gaps. None of them is solved by a bigger model alone.

The vocabulary, made precise.

The terms below were sharpened in recent design work. They are deliberately distinct: each names one thing and only that thing. Two glosses anchor the whole model: lenses are about what data exists for me; frames are about what data is relevant right now.

Substrate Corpusdurable · one per principal
The durable, append-friendly store of all signed facts and content artifacts (any medium, addressed by content hash), plus state-changes, decisions, and provenance. There is one corpus per principal. It lives in persistent storage: SQLite on a phone, Postgres in a VPC.Where all the data lives. The cold authority.
Substrate Referencesigned · DID-addressable
A signed, DID-addressable pointer to a corpus: a handle that lets a party query without holding a copy. Cross-organization operations exchange these references, so the data stays with its owner.
Lensconsumer-scoped · flat · additive
A consumer-scoped read filter. Flat and additive: wearing more lenses widens scope. There is no hierarchy and there are no tiers.Lenses are about what data exists for me.
Frametask-scoped · transient · narrowing
Task-scoped, transient, narrowing: the picture frame around the work in focus. Frame is the universal primitive: a "session" is just one frame_type. A clinician's chart-open, a customer's "at the shop," and a deal-open are others.Frames are about what data is relevant right now.
Substrate Slicein-RAM · lens-projected · frame-bound
The in-RAM, lens-projected, frame-bound materialization an agent holds during a session: the product of applying a lens and a frame to the corpus. (Formerly called a "view.")The warm cache. The corpus is the cold authority.
render(corpus, lens, frame) → slice
Read it as: applying a lens and a frame to the durable corpus renders the transient slice the agent works from. Time-scoping lives in the frame, so there is no separate "window" parameter.
Corpus
cold authority · durable
Lens
what exists for me
+
Frame
relevant right now
render()
project + bind
Slice
warm cache · in-RAM

render(corpus, lens, frame) → slice. The corpus is the authority; the slice is what the agent holds.

Files vs. the database.

A second design split runs underneath the model: files are the agent's engagement surface; the database is the navigation index.

Files: the engagement surface

Curated markdown with wiki-links

Agents traverse curated markdown by following wiki-links, the way a person reads, one connected note to the next. This is where the agent actually works.

Database: the navigation index

Answers cross-cutting queries

The database answers cross-cutting queries and points back to the files. Embeddings are the discovery complement, a way to surface what to read.

The protocol, and its reference implementation.

SLF (Substrate · Lens · Frame) is the protocol and architectural layer. Alexandria is Lexenne's reference implementation of it. SLF is a governance protocol for AI memory: it decides who sees what fact, in what role, with what audit trail.

SQLite is to SQL as Alexandria is to SLF: the reference implementation, free and exemplary. Any conformant implementation works.

At a high, conceptual level, SLF is built around a few commitments:

Intrinsic per-fact gates

Every fact carries its own access rules and regulatory metadata. Governance travels with the data itself.

Signed audit receipts

Actions emit a verifiable receipt as a first-class output. The receipt is part of the action itself.

Interoperability & sovereignty

The reference implementation is Apache-2.0. The design is survivable without Lexenne: if the company disappears, a conformant implementation still works and the principal still owns their corpus.

Filling the contextualization gap

SLF is designed to fill the "contextualization gap" that large-memory-language-model research acknowledges but does not solve: deciding who sees what, in what role, with what trail.

This page covers what SLF is and why it matters. The wire-level operations, the gate and receipt mechanics, the grant and identity machinery, and the governance specifics live in the design documents, deliberately out of scope here.

Status: SLF is a v0.4 draft design. Reference implementation is about to begin.

The hard part is deciding what to keep.

Storing and retrieving is the easy part. The hard part, the part that is the moat, is deciding what to keep: reweaving new knowledge into old, resolving contradictions when two facts disagree, and forgetting what has gone stale.

Conflict resolution is a frontier problem the broader field has not solved. It is the moat. The mission and the hardest technical problem turn out to be the same problem.

Lens-filtered cross-session retrieval is live: an agent starting a session can pull in relevant prior context. The compounding loop itself, reweaving, conflict resolution, and forgetting, is being built now.

One shared substrate that every tool reads from.

The common approach is per-tool memory: each product ships its own memory feature, each one starts over, and none of them talk to each other. The knowledge fragments across silos that each begin from zero.

The alternative is one shared substrate that remembers state-changes across systems, read through role-shaped lenses. One memory, many lenses.

The durable position is the unified substrate. Anyone can add memory to a single tool. The hard, defensible thing is one corpus that compounds across all of them.

Current state: what is real, in build, and designed.

Figures grow over time. Three status levels: REAL (running today), IN BUILD, DESIGNED. "Running" here means running on a personal development fleet.

Real
Separated substrate corpus. Exists as a separated store (~26 tables); ~2,250 lessons and ~2,798 decision traces captured; 13 consumers; 7 lens contracts.
Real
Lens-filtered, token-budgeted cross-session retrieval. Live. The same window, read by a different consumer, yields different rendered content.
Real
Session lifecycle. Register / load-context / work / end-with-reflection; decision traces; a session digest (reduce, reflect, archive) that runs post-session.
Real
The agent-development framework (Forge) that dogfoods all of this.
In Build
The compounding loop. Reweaving, conflict resolution, forgetting (pruning, supersession, stale-detection), capture-deduplication, and frames as a live surface.
In Build
Serendipitous retrieval. An agent-pull substrate query: the agent reaches into the corpus on its own to find what it needs.
Designed
SLF, the Substrate · Lens · Frame protocol. A v0.4 draft. Reference implementation is about to begin.
Partial · building
Lexenne Studio + Patina products. Subdomains provisioned; apps not yet shipped.