docs: brain-redesign planning dossiers (pre-grilling)

[thinmintdev]

What

Nine cross-linked planning documents in docs/internal/brain-redesign/ for a unified-memory "brain" redesign of the hal0 platform: a two-tier system pairing a semantic/episodic engine (Hindsight candidate, replacing Cognee) with a human-legible compiled-knowledge wiki (Hal0ai/hal0-wiki / Karpathy "LLM Wiki" pattern), integrated across every hal0 agent/app/MCP surface and fully self-hosted on Strix Halo.

Doc	Purpose
00-INDEX	Front door — TL;DR, reading guide, exec summary, glossary
01-hindsight-research	Exhaustive Hindsight feature inventory (banks, bank-templates, retain/recall/reflect, webhooks, ops, MCP, integrations)
02-hal0-wiki-research	The Obsidian LLM-Wiki framework (skills, vault model, Hermes/Pi/Claude install)
03-current-memory-audit	As-built Cognee subsystem with file:line citations
04-consumer-surface	Every brain consumer + the deep-integration hook map
05-memory-landscape	Paradigm taxonomy, tradeoff matrix, two-tier thesis
06-proposed-architecture	The recommendation: layers, MemoryProvider ABC, promotion model
07-integration-roadmap	7 ship-able phases, dependency graph, eval hard-gate
08-grilling-decision-tree	Working doc for the grilling session — dependency-ordered decisions, each with a recommended answer + counter-challenge

Why now

• Cognee's graph layer is dead weight — off by default because local <32B models can't do reliable triple extraction; hal0 effectively runs "pgvector + a fragile SQLite filter sidecar."
• No staleness handling today; Hindsight's consolidation/observations directly addresses it.
• memory_add normalizes dataset to "shared" — private:* namespaces silently flattened #317 (private-namespace no-op) is already fixed (PR fix(memory): REST /api/memory/add honors private:* namespace #366/fix(memory): honor per-call client_id on read path + audit (Phase D regression) #369) — not a driver; flagged so it isn't relitigated.
• Hindsight runs fully local (embedded Postgres + BGE + MiniLM rerank + built-in LLM, or route extraction to lemond:13305); Strix-Halo accel is undocumented → plan CPU embed/rerank via existing slots.

Status

Planning only — nothing decided or implemented. Recommendations are opening positions to be stress-tested in a /grill-me session (start at doc 08). Multi-agent research wave, 2026-06-02. Targets v0.4. Will supersede ADR-0005/0014 once decided.

🤖 Generated with Claude Code