📝MemPalace, 📝Obsidian, and 📝MythOS are three different answers to the same question: how does an AI collaborator carry knowledge across sessions? They sit at three distinct points on the spectrum from raw capture to curated artifact. MemPalace stores conversation history verbatim and searches it semantically. Obsidian is a file-system of user-authored markdown notes that AI clients bolt onto via 📝MCP. MythOS is an AI-native knowledge platform with structured memos, identity-aware augmentation, permissioned retrieval, and a social layer. This memo extends the prior pairwise comparisons — 📝MemPalace vs MythOS and 📝Obsidian vs. MythOS as Claude Memory — into a three-way view.

The Three Models

MemPalace — verbatim memory. A local-first Python system on ChromaDB with a SQLite temporal knowledge graph. It mines conversation logs (Claude Code JSONL, ChatGPT, Slack, plain text) and stores chunks verbatim, organized by a Method-of-Loci metaphor (Wings, Rooms, Halls, Drawers). Retrieval is vector search with ~170-token tiered context loading. The unit of memory is a raw conversation fragment, never read directly by humans.

Obsidian — file-system notes. A mature desktop note-taking app with 1,800+ plugins and a local-first markdown vault. AI memory emerges when users point an MCP server at the vault and drop a CLAUDE.md in the root. The unit of memory is a user-authored markdown file, read by both the human and the AI. Semantic search arrives by plugin — Smart Connections embeds the vault on-device — but most Obsidian MCP servers still dump whole files into the context window.

MythOS — curated AI-native knowledge. A web platform with structured memos (Objective/Subjective/Contexts), an Augmentation System (Soul, Style, Human, Memory), three-tier visibility, @mention knowledge graphs, and tiered context loading (L0/L1/L2). Accessible via web, mobile (Claude.ai OAuth), 📝Claude Code skill, MCP, and email. The unit of memory is a curated artifact — human-readable, shareable, permissioned.

Capture Model

MemPalace captures after the fact — it mines session logs post-hoc, so the AI has no influence over what gets stored. Obsidian captures when the user writes a note, or when the AI writes into the vault at the user's instruction. MythOS captures live during conversations via MCP — the AI recognizes memo-worthy moments and creates structured artifacts in the moment. Live-during-conversation capture is higher-signal than post-hoc log mining, because the AI is a present participant in deciding what's worth keeping.

Storage Primitive

Identity Layer

MemPalace has a ~100-token plaintext identity file. Obsidian has CLAUDE.md — a manually-maintained file that drifts stale because nothing structurally keeps it current. MythOS has four living augmentation memos (Soul, Style, Human, Memory) auto-loaded before every operation via get_context, evolved through ongoing collaboration. Identity-aware augmentation is a MythOS-native capability; the other two require the user to hand-maintain it.

Access Control

MemPalace is single-user and local — access control is whoever has the machine. Obsidian MCP exposes the entire vault to any query, which surfaces real privacy concerns when personal journals sit beside professional notes. MythOS has three-tier visibility (public, link, private) plus audience tags that flow through the RAG pipeline, so a memo tagged #for-close-friends is invisible to unauthorized retrieval.

Platform Reach

MemPalace runs wherever Python runs — MCP and CLI, desktop-oriented. Obsidian's MCP integration needs filesystem access, which means the desktop app running or a vault kept alive on a server you administer via the official headless Sync client; either way Claude mobile and web reach it only through a machine you maintain. MythOS works across web, iOS, Android, Claude Code, and any MCP-compatible client via OAuth 2.0 and an npm package, with no machine of yours in the loop. Mobile continuity — starting a session on a laptop and continuing on a phone against the same memory — needs nothing running but MythOS.

Social & Collaboration

MemPalace and Obsidian are both single-user by design. MythOS has communities, forking, public memos, newsletters, and shared knowledge graphs — knowledge built in one library can flow to others under explicit permissions.

Context Loading

MemPalace pioneered tiered loading — ~170 tokens at wake-up, room-level context on demand, leaving 99%+ of the window free. Most Obsidian MCP servers dump whole files, which scales badly past a few hundred notes; the embedding plugins help inside the app but don't change what the MCP server hands the model. MythOS adopted MemPalace-style tiering in April 2026: get_context returns L0 (Soul + Style) + L1 (Human + Memory), and load_context_tier pulls L2 (additional context memos + template catalog) on demand.

Portability

MemPalace is a local filesystem — fully portable by definition. Obsidian is also local markdown — portable, but only within markdown-reading tools. MythOS ships bidirectional portability: Markdown export (npx mythos-mcp pull) produces .md files with YAML frontmatter compatible with Obsidian, Hugo, or any markdown tool, and Obsidian Import (npx mythos-mcp import --obsidian ~/vault) converts [[wikilinks]] to @mentions and uploads images. Lock-in is answered in both directions.

Where Each Wins

MemPalace wins when the use case is truly verbatim — a power user who wants every word of every conversation retrievable, with temporal validity on every claim, running entirely local with no cloud touchpoint. It is the best raw-memory engine in the open-source ecosystem.

Obsidian wins on price (free, commercial use included), ecosystem depth (1,800+ plugins), zero-friction onboarding ("just open a folder"), and offline-everywhere. For users who already live in Obsidian and want a light AI memory bolt-on, the switching cost outweighs the gains.

MythOS wins when the goal is AI-native collaboration — identity-aware augmentation, permission-aware retrieval, cross-platform continuity (web + mobile + desktop + MCP), curated human-readable artifacts, and a social layer. It is the only one of the three designed for AI collaboration from day one rather than adapted into the role.

FAQ

Can I use more than one?

Yes. All three expose MCP servers, so Claude can call into MemPalace, Obsidian, and MythOS in the same conversation. They answer different questions and do not conflict — MemPalace gives you "what did we say?", Obsidian gives you "what did I write?", MythOS gives you "what do I know, and who am I?"

Which one is most beginner-friendly?

Obsidian — open a folder, start writing. MythOS is next: account signup, browser-based editor. MemPalace requires Python and ChromaDB comfort, plus MCP configuration.

Which one scales best past 1,000 memories?

MythOS and MemPalace both. MythOS uses 📝RAG with vector embeddings; MemPalace uses ChromaDB directly. Obsidian's MCP servers typically degrade past a few hundred notes because they dump files into context.

Which one is best for cross-device use?

MythOS — web, mobile apps, CLI, and any MCP-compatible client via OAuth or API key. MemPalace is local-first; Obsidian's MCP integration needs filesystem access, so reaching it from a phone means keeping a desktop or a headless-sync server of your own running.

Which one is best for team or shared knowledge?

MythOS — three-tier visibility, communities, and a social layer. MemPalace is fundamentally single-user. Obsidian has external add-ons (Publish, Sync) but they were not built for AI collaboration.

Future State

MythOS is absorbing the strongest engineering patterns from MemPalace (tiered loading shipped; context cost transparency shipped; temporal validity on relationships deferred) and closing the Obsidian onboarding gap through Markdown export and Obsidian Import. The strategic position: a curated, sovereign, permissioned knowledge system that matches MemPalace's memory engineering and Obsidian's portability while preserving its own advantages — identity, social layer, structured artifacts, cross-platform reach.

These three tools aren't really competitors — they're three different theories of what "AI memory" means. MemPalace says memory is the raw tape. Obsidian says memory is the notes you happen to have already written. MythOS says memory is a curated, structured, permissioned artifact co-authored by the human and the AI. The MythOS thesis is that curation at the point of capture — with human intelligence as the extraction layer, not an LLM and not post-hoc log mining — produces higher-signal knowledge than either alternative. MemPalace validates this from the engineering side (verbatim beats LLM extraction) and Obsidian validates it from the market side (people want persistent AI memory badly enough to hack it together with CLAUDE.md and MCP servers). Both point at the same conclusion from different angles.

📝MemPalace vs Obsidian — the pairwise comparison of the two non-MythOS options, for readers evaluating those two alone
📝MCP vs RAG — the retrieval-strategy lens underlying these architectures
📝Karpathy's Flag in the Ground — LLM Knowledge Bases — the "LLM Knowledge Base" pattern that frames this category
📝The Augmentation Stack — MythOS's framework for memory, mind, and mouth

Brian Swichkow built MythOS because neither a verbatim backend nor a Markdown vault gave him what he actually needed: an AI that knows who he is, works from his structured knowledge, respects his permission model, and follows him across every client and device without re-explaining anything. MemPalace and Obsidian are excellent tools, recommended without hesitation for the problems they solve best — they were never going to be MythOS because they are solving different problems. The right question for a given builder is not "which is best?" but "which layer am I missing?" Capture, curation, or augmentation — the answer tells you which one to install first.