Stage 1 — Dev-Scoped "Second Brain" (RAG over your own context)

New topic, not covered by Topic #10 (memory-and-progress-ledgers.md, which is about a dev/project progress ledger, not a full retrieval system over your own codebase/notes/history). Scoped per explicit user confirmation to the developer use case — querying your own codebase, ADRs, past troubleshooting, personal snippets — not general life organization. No pick made; options and trade-offs only.

1. OpenClaw — name-collision resolution (the important part)

This is not a naming collision. It is the same project, and the security incident documented in security-and-supply-chain.md is directly relevant, not a false alarm.

Not a naming collision: the same project as the security incident.

OpenClaw (github.com/openclaw/openclaw) is real, and large: a local-first AI gateway that runs on your own machine and talks to you through whatever chat app you already use (WhatsApp, Telegram, Signal, etc.).

Origin: Peter Steinberger (the developer behind the PDF SDK used in Dropbox/Slack/Box) pushed it to GitHub in November 2025 as "Warelay," renamed "Moltbot," then "OpenClaw" after a trademark dispute.
Scale: one of the fastest-growing open-source projects on record — 247,000 stars, 47,700 forks by March 2026.
Steinberger joined OpenAI in February 2026. The project moved to an independent foundation, stayed open, later shipped native iOS/Android apps.

Its extensibility model is "claws" — installable plugins distributed through ClawHub. That's the same skill registry security-and-supply-chain.md documents as systematically poisoned in the February 2026 campaign (30+ malicious skills at peak infection, SlowMist flagging 472; Snyk's ToxicSkills scan found 36.8% of scanned ClawHub/skills.sh skills had a security flaw, 13.4% critical). Not two different "OpenClaw"s — the core project's own plugin marketplace got compromised, for the same structural reason already noted: publishing to ClawHub needs only a SKILL.md and a week-old account, no code signing or default sandboxing.

Practical read: OpenClaw itself (core gateway, self-hosted, no third-party claws installed) is real, actively maintained, legitimate — not a scam or a hallucination. The risk is specifically in what you install from ClawHub. Treat any third-party "claw" like any other unaudited MCP server or skill, per the vetting checklist in security-and-supply-chain.md — don't install broadly just because the core project is huge and legitimate.

2. "Open Human" — resolved, real project under a different exact name

The claim ("Rust-based personal super-intelligence, encrypted memory tree over docs/emails/meetings, MCP integration") maps to a real project: OpenHuman (github.com/tinyhumansai/openhuman) — not "Open Human," a separate name.

Right project, slightly wrong name. Primary docs exist under the correct one.

Built in Rust + Tauri, GPL3-licensed, maintained by Tiny Humans AI.
Summarizes/compresses documents, emails, chats into a "memory graph."
Supports on-device encryption, an approval gate, OS-keyring secrets, opt-in sandboxing, a "Privacy Mode" that keeps all inference local.
Exposes memory over MCP, and can drive MCP servers itself (including OAuth flows).

Real, checkable maintainer org, active project. The original dump's sourcing (a single YouTube video) undersold something that has real primary documentation once you have the correct name.

3. Established RAG-over-personal-context architecture (consensus, not a new claim)

The pattern in the pasted dump is standard, current, and settled enough to state as consensus rather than re-verify piece by piece:

Checked and real. The dashed card is a list, not an authority.

Reuse Postgres, run a dedicated store, or take the simplest local option.

The model never searches your disk. Retrieval pastes the best matches into its context first.

[ Ingestion ]              [ Vector store ]           [ LLM / interface ]
Git commits, markdown  →   pgvector (Postgres) /   →   Ollama (local model)
notes, exported Slack/     Qdrant / ChromaDB            + Open WebUI / LibreChat
Jira, code repos            (via LlamaIndex/LangChain)   + IDE extension (Continue.dev)

Ingestion/chunking layer: LlamaIndex (run-llama/llama_index, already in _ai-tooling-recommendations.md) or LangChain parse and chunk source material — code, markdown, exported chat/ticket logs — into embeddings.
Vector store: pgvector (a Postgres extension — convenient if you already run Postgres, avoids a separate service), Qdrant, or ChromaDB (simplest to run locally, SQLite-adjacent ergonomics) are the three real, well-established options. None is a wrong choice; the difference is mostly "do you want another service running" vs. "reuse Postgres" vs. "simplest local file-backed option."
Interface layer: Open WebUI (open-webui/open-webui, already in _ai-tooling-recommendations.md's browser-tooling section for a different reason — it's also a legitimate offline chat front-end) or LibreChat for a chat-style interface; Continue.dev for pulling that same custom context directly into VS Code/JetBrains as inline completions/chat, which is the more natural fit if the goal is staying in the editor rather than context-switching to a separate chat window.
Model: Ollama running a local model — directly continuous with the hardware analysis already done in local-model-hardware-fit.md (this rig: RTX 3070, 8GB VRAM, 64GB system RAM — full-VRAM ceiling around 9B, larger models RAM-offloaded and slower).

Other named projects from the dump, checked and real: Dify (visual RAG-pipeline/app builder), Logseq (logseq/logseq — privacy-first outliner/knowledge graph with an AI plugin ecosystem), RAGFlow (infiniflow/ragflow — deep document-understanding engine for messy unstructured files). aristoapp/awesome-second-brain is a real curated list, general-scope (not dev-specific) — useful as a browsing entry point, not a citable authority in itself.

4. Practical read for this setup specifically

Per this repo's own anti-speculative-infrastructure stance (docs/reasoning.md: extract from real need, don't build ahead of it) and the hardware already analyzed in local-model-hardware-fit.md:

Extend work already in progress instead of new infrastructure.

Worth trying now: Continue.dev + Ollama, pointed at just this repo's docs/decisions/planning files — not a general RAG stack yet. Extends work already in progress here instead of standing up new infrastructure.
Premature right now: a full ingestion pipeline (Slack/Jira export → vector DB → dedicated UI) for a solo, pre-product-code repo. Not enough accumulated history yet to make retrieval meaningfully better than reading the files directly.
Skip entirely for now: OpenClaw and OpenHuman as personal-assistant-with-messaging-integration products — they solve a broader "AI across your whole digital life" problem, not the narrower "RAG over my own codebase" one this stage was scoped to. Worth knowing they exist (OpenHuman's MCP-native memory-graph approach is architecturally close to Topic #10), not worth adopting yet.

Open questions / weak sourcing

OpenClaw's current ClawHub security posture (whether default sandboxing or code signing has since been added, given how fast the project is moving) wasn't re-checked here — security-and-supply-chain.md's Feb 2026 findings should be treated as a point-in-time snapshot, not necessarily still true today.
aristoapp/awesome-second-brain's star/fork counts and activity level weren't independently pulled — confirmed real via search, not repo-health-verified to this project's usual bar.
Continue.dev's current feature set for pointing at a custom local vector store (vs. its more common use with hosted/cloud context providers) wasn't verified against its current docs — worth a direct look before actually wiring this up.

Unverified. Look before relying on any of these.