Compounding knowledge across people and agents, so every practitioner gets the advantage.
A forward-looking vision of where Alexandria is headed.
Today, the best evidence lives in people's heads. When they leave, it leaves. When agents start a new session, they start from zero. The organization never compounds.
Individuals still get the serial practitioner's advantage. The difference: so does everyone else, and every agent, every session.
Capture is already solved: organizations generate knowledge constantly. The problem is connection, maintenance, and retrieval at the moment of need.
Substrate-vs-lens is a proven bet. The same architectural shape has run in production at Macmillan Learning under three different names. Alexandria is its solo-scale continuation.
UXF (Customer Experience Contextualization Framework): one customer-experience substrate, multiple role-shaped lenses. ASO26 / IA-EA: per-course persistent record (course + institution + time) was the substrate; SFDC decision-trace objects with as-is vs as-was fields were state-change as first-class; sales / support / leadership dashboards were the lenses. EDR > Decision Support: role-conditioned surfaces (CEO vs sales vs support) over the same memory. The pattern compiled, shipped, and ran.
| Macmillan artifact | Substrate / Lens role |
|---|---|
| UXF customer-experience contextualization | Customer-facing lens over substrate |
| ASO26 per-course persistent record | Bounded entity timeline (substrate) |
| ASO26 decision-trace SFDC objects (as-is vs as-was) | State-change primitive in microcosm |
| ASO26 Profile vs Context (two-tier identity) | Lens (Profile) vs Frame (Context) |
| EDR > Decision Support role-conditioned surfaces | Role-conditioned lenses over same memory |
Decision-trace and as-is/as-was state-change machinery shipped at Macmillan years before Sentra named the substrate-vs-lens pattern. Alexandria carries that lineage forward at solo scale, with the architectural shape that admits enterprise lens packs as additive, layered on without a rewrite. Two 2026 empirical papers supply external validation of the approach: Cao et al. (arXiv 2603.20432) show +17.3% from agents with filesystem-organized context; Lee et al. / Stanford (arXiv 2603.28052) show +7.7 points with 4× fewer tokens. The architecture we already built is exactly what both papers measure.
Alexandria is a knowledge architecture, distinct from a wiki, a search engine, or a chatbot. It's the system that makes knowledge compound across people and agents: shared, connected, and maintained so every team member and every agent session gets the evidence base that used to live only in the veteran's head.
The universal state of the organization: entities, facts, state-changes, decision traces, interactions, bi-temporal validity. Provenance, permissions, history. One substrate per organization; everyone shares it.
The role-shaped filter. Sales sees opportunity health. Engineering sees code precedent. The CEO sees strategy drift. Same substrate, different surface, governed by RBAC and preferences. Lenses are how the substrate becomes useful per consumer.
The transient task binding. What is this person, agent, or team currently working on, what's the plan, what's been tried, what's next? Frames hold the work in front of you; lenses hold who you are.
Substrate · Lens · Frame is also a protocol, SLF, that Alexandria implements as its reference. The pattern is built to be portable across implementations. The SLF design →
Every action feeds back. Every retrieval is a usage signal. Every correction strengthens the graph for the next person and the next agent. By design, the loop compounds across the organization.
Humans do what they're already doing: talk to customers, make decisions, write a note, lead a meeting. Agents do what everyone intends but no one has the time or cognitive bandwidth to do at scale: connect, track, resurface, detect drift, find patterns across fifty conversations, so the organization compounds alongside the individual.
Low friction in. High value out. Capture is cheap, a byproduct of work you're already doing. Compounding is shared: every correction, every connection, every lesson improves the graph for the next person and the next agent. The thing that killed every wiki, the maintenance no one gets to, is what agents are built for. They don't get bored. They don't skip the update. They notice the contradiction you forgot about three months ago.
Each lens is a contract: entity types loaded by default, relationships in scope, state-changes watched, RBAC capabilities, expiration policy. The substrate generalizes; lenses don't. Sales, Success, On-call, Exec, and Customer-facing lenses are future lens packs on the same substrate. Lens packs are the productizable surface; substrate is the platform.
| Lens | Default scope | Watched state-changes |
|---|---|---|
| Architect | ADRs, planning docs, system design, integration points | Architectural drift, ADR contradictions, cross-system coupling |
| Developer | Code, tests, gates, eval results, golden principles | Test failures, regression signals, gate breaks |
| Analyst | Requirements, decision matrices, stakeholder context | Stated-vs-revealed preference drift, assumption invalidation |
| Steward | Backlog, sprint state, dependencies, claim leases | Stale claims, ghost tickets, blocked dependencies |
| Writer | Docs, Confluence, public-facing surfaces | Stale docs, undocumented features, drift from code |
| Principal | Workspace-wide; can wear any lens | Strategy-execution drift, longitudinal patterns, cross-entity reconciliation |
| Subsystem | Trigger-shaped, event-conditioned | Threshold crossings; emit act/wait/escalate/no-op decisions |
| Per-tool memory (Notion / Linear / Cursor) |
Substrate (Alexandria) | |
|---|---|---|
| State of the world | Fragmented per tool | Unified in one substrate |
| State change | Audit log per tool | Typed state-change with before/after, consumer, frame, evidence |
| Time | created_at only | Bi-temporal (valid_at + invalid_at) |
| Personalization | Tool is the lens | Persistent lens contract with RBAC capabilities |
| Task context | Open tabs and chat history | Frame: intent + next-steps + footprint + suppression |
| Cross-tool reasoning | Glean searches; nothing preserves change | Substrate is queryable across every system that writes to it |
Memory / Learning / Personalization is the decomposition inside the substrate side of the architecture. Memory stores, Learning processes, Personalization serves. The lens layer sits above; the frame layer sits above the lens. view = render(substrate, lens, frame).
Notion remembers Notion-stuff. Linear remembers Linear-stuff. Cursor remembers Cursor-stuff. Glean searches across them but doesn't preserve state-change. Sentra is building substrate + lens for enterprise. Per-tool memory is the failure mode the industry is collectively producing.
The credibility artifact is years of solo dogfood under load, grounded in shipped work. ASO26 proved the pattern shipped. Alexandria is proving it compounds longitudinally. Lens-pack productization, if it ever happens, follows from those two proofs.
Every assumption explicit, with kill criteria and a validation timeline. When an assumption dies, the system knows what decisions depended on it.
Separate the quality of a decision from the quality of its outcome. Retrieve "well-reasoned precedent", not just "what worked last time."
Compare stated priorities against behavioral evidence: time allocation, meeting topics, actual decisions. Surface drift before the quarterly review discovers it.
Patterns emerge at the intersection. Customer signals connect to product assumptions. Decision traces link to strategy claims. The graph sees what individuals can't.
No single metric triggers action. The signal is multiple independent indicators crossing thresholds in a time window. The agent watches for clusters rather than isolated spikes.
Not a data dump. Retrieval filters → full claims → wiki-links → traversal. The agent exercises judgment at every layer about what to absorb.
Universal storage of state: entities, facts, typed state-changes, decision traces, interactions, bi-temporal validity. Semantic-filesystem-shaped, distinct from graph-DB-shaped. One substrate per organization; everyone shares it.
Persistent, consumer-keyed, DB-resident filter. Declares entity types loaded, relationships in scope, watched state-changes, RBAC capabilities, expiration policy. Lens mutations flow through the same approval-gated pipeline as substrate mutations.
Transient, task-keyed, DB-resident binding. Holds canonical intent (one of ten), next-steps with assigned consumers, footprint of decisions made, surfaced substrate references, suppression state. Scratchpads carry frame_id at creation.
Knowledge that compounds across people and agents, accumulated, connected, challenged, and maintained, so everyone gets the evidence base that used to live only in the veteran's head.
Your only job is to capture. The system compounds for everyone.