Autri — Platform Vision (north-star seed)#

Seed captured 2026-06-01 from a Sunday→Monday brainstorm. NOT a committed design — beta comes first. Captured so it isn't lost. Companion to memory project_autri_platform_thesis and decisions.md D2/D55.

The one-liner#

Autri takes whatever AI you're using and makes it better — without training on new data. It is the inspectable knowledge substrate; agents are thin task-layers on top, connected via MCP.

Thread 1 — Self-documenting dev memory (dev tool: Dan + Claude)#

A /hl:stop hook captures the Claude Code session transcript → an LLM segments it into decision-episodes (problem → options → decision → rationale → files touched) → each embeds as a chunk → MCP search retrieves them.

• Three layers: autri = long-term provenance · three-file system (methodology/next/decisions) = working-set contract · Foundry = published design surface.
• Does NOT replace Foundry (curated/normative). It feeds it and cheapens /hl:stop curation (LLM proposes the decisions.md/annotation diff, human approves).
• Hard design rule: wire RETRIEVAL into /hl:start, not just capture at /stop — else it's a write-only journal.
• Risk: sloppy episode-extraction carries false authority; distillation quality is the whole ballgame (LLM-does-semantics, verify).
• Why it'd help the agent (Claude's own read): kills cold-start re-discovery (cost ~half a session on the "New chat" bug that was mis-recorded as open), enables decision archaeology + cross-session pattern extraction (the ~30 hand-written memory files are a manual version of this).
• A broader consumer north-star exists (memory-over-MCP for any assistant) but consent + signal-to-noise make "index everything wholesale" the wrong v1.

Thread 2 — Substrate for knowledge-grounded agents#

QuoteAI = a quoting agent on a KB. Generalizes to proposals / contracts / reports / RFP responses / research / support.

• Bound the wedge: knowledge-grounded agents (competence from a trusted, retrievable, inspectable corpus), NOT "any agent for any task" (that's orchestration — a different problem; don't boil the ocean).
• QuoteAI-on-autri = the first vertical / external proof point (de-risks productization with a real Brehob customer). It consumes autri as an MCP/API service, not a code merge (see D2, 2026-06-01).
• Immediate low-risk dogfood: index Foundry's own design docs into autri so MCP search over project knowledge is great — proves retrieval quality before trusting it with conversation-episodes.

Enterprise tenancy (when the commerce tier opens)#

Pool (row-level, D13) for self-serve; silo (own VPC/RDS/S3, ideally own AWS account) for enterprise. Brehob = silo #1. Enabler: parameterized CDK so a silo is config+deploy, not a fork. Bridge model (DB-per-tenant in a shared cluster) is the cheaper middle rung. Isolation knobs (per-tenant KMS, bucket prefixes) layer onto any rung.

Next step when we're ready#

Prototype Thread 1's /stop capture + episode extractor as a parallel experiment (manual-first; don't make the methodology depend on beta-stage autri). Measure retrieval quality on Foundry-docs-in-autri first.