Skip to content

Latest commit

 

History

History
78 lines (53 loc) · 6.11 KB

File metadata and controls

78 lines (53 loc) · 6.11 KB
status EMERGING
last-verified 2026-04-28
evidence-tier C
applies-to-signals
memory-systems
cross-project-synchronization
federation
revalidate-by 2026-10-28

Archetype D — Cross-Project Portfolio Brain

Evidence Tier: C — federation is hand-rolled today; the OpenBrain compilation-agent path is Tier D (not shipped).

Purpose

Per-archetype recommendation for managing multiple knowledge-bearing repos as a portfolio — preserving per-repo evidence and provenance while still surfacing cross-repo connections.

D1. Primary stack

Per-repo (LLM Wiki + Graphify) + a thin federation index. Federation today is hand-rolled. OpenBrain's compilation agent is roadmap, not shipping.

Layer Owner Why
Per-repo Karpathy wiki + Graphify (as in archetype A) Preserves per-repo evidence tier and provenance
Federation Small ~/portfolio/INDEX.md + nightly cron merging per-repo index.md files Axis 4 — federation, not consolidation
Cross-repo topology Second graphify run over all repos → cross-repo communities tagged INFERRED Axis 8 — separates within-repo (EXTRACTED) from cross-repo (INFERRED)

Driving axes: 4 (cross-tool/cross-repo), 2 (augments per repo), 8 (provenance discipline survives federation). Tier: C; OpenBrain compilation-agent path is Tier D (not shipped).

D2. Hybrid alternatives

Hybrid Optimizes Pick when
Per-repo + OpenBrain federation (when compilation agent ships) Live cross-repo recall FSL-1.1-MIT terms acceptable; want concurrent multi-tool reads
Per-repo + claude-context only on code-heavy repos Semantic recall in code without forcing every repo into vectors Some repos are code, some are prose
Single mega-repo with tagged subfolders Drops federation complexity Repos are small and you're really managing one knowledge body

D3. Anti-patterns

  • Single-DB consolidation across all repos (OpenBrain Postgres, claude-context Milvus): flattens per-repo evidence tiers; one repo's schema decisions infect another's. A single Milvus collection with vectors from prose + AST chunks produces nonsense neighbors.
  • Shared cross-repo graph as the only graph: communities bleed across project boundaries. The cross-repo graph must be a secondary INFERRED-only view.

D4. Adoption order

  1. Each repo independently passes archetype-A adoption first.
  2. Write a 50-line script concatenating per-repo index.md files into a portfolio INDEX.
  3. Run graphify ~/portfolio over all repos to see what cross-repo edges emerge. Stop if edges are mostly noise.
  4. Re-evaluate OpenBrain only after the compilation agent ships and you've reproduced its claims.

D5. Constraint check

All six met when run as per-repo + thin federation. OpenBrain path defers local-first to self-hosted Postgres setup; FSL terms must be accepted.

Sources

Inherits source rubric and tier methodology from memory-systems-recommendation-methodology.md.

This variant carries no new external sources with independent URLs beyond what the methodology doc and Archetype A already cite. The per-repo layer inherits Karpathy (Tier B paradigm) and graphify (Tier C, safishamsi/graphify) from Archetype A. The OpenBrain compilation-agent path is explicitly labeled Tier D (not shipped) in the doc body.

Related Analysis

Related (from graph)