Use Obsidian as the default second-brain layer around Hermes, with Mem0 as an optional semantic-recall sidecar. Hermes's built-in MEMORY.md and USER.md should remain the primary curated memory lane. Obsidian is the better human-owned canonical knowledge base for real personal workflow: durable local Markdown, inspectable notes, links, backlinks, long-form thinking, and low lock-in. Mem0 is better at AI-native fuzzy recall over interaction history, but it should be additive retrieval infrastructure, not the place Brian trusts as the canonical source of truth.
Hermes already has a deliberate split: curated native memory (MEMORY.md and USER.md) is the prompt-injected primary lane; external memory providers such as Mem0 are secondary/additive, queried explicitly for semantic recall. That architecture strongly favors not treating Mem0 as a full replacement for Hermes memory. Mem0 fits as a provider-lane tool: “find candidate memories, then promote durable truths into native memory or a skill.” Obsidian fits a different layer: a human-readable corpus that Hermes can search, read, edit, and cite when the user wants durable notes rather than agent preference memory.
Tradeoff: Mem0 is closer to Hermes's agent loop; Obsidian is closer to the user's brain and archive.
Obsidian wins as canonical personal knowledge. A vault is a folder on the local filesystem. Notes are Markdown, plus open-ish adjacent artifacts like Canvas, PDFs, images, and media. A user can grep it, back it up, sync it, edit it, and read it without an AI system.
Mem0 wins as secondary recall. Its API and product model are purpose-built for add/search/update/delete memory operations, semantic search, entity scoping, events, and agent memory. That is exactly what a semantic sidecar should do. But because memories are extracted, embedded, ranked, and sometimes accumulated from conversations, Mem0 is less suitable as the authoritative place to reason, revise, and settle meaning.
Obsidian's core premise is local-first: your vault lives as files on your device, no account required, and Obsidian says no telemetry is collected. Obsidian Sync is optional and defaults to end-to-end encryption for remote vaults.
Mem0 can be self-hosted and the OSS docs explicitly say you own the stack, data, and customizations. That is much better than a purely hosted memory API. But it still introduces more infrastructure: LLM provider, embedding model, vector store, optional Postgres/pgvector or Qdrant, auth, dashboard, API keys, background processing, and service health. For a personal second brain, more moving parts means more ways memory silently degrades.
Mem0 is the clear winner here. Its current docs describe memory CRUD, semantic search with filters, entities, events, async writes, hybrid retrieval, entity matching, BM25, temporal reasoning, and memory-specific benchmarks. The arXiv paper reports better long-term conversational recall than multiple baselines, lower latency and lower token cost than full-context approaches.
Obsidian's native strengths are search, backlinks, graph view, tags, and files. It can become semantically searchable through plugins, local embeddings, MCP/file tools, or a custom index, but that is not the product's default center of gravity. Obsidian is a knowledge workspace first; Mem0 is a memory retrieval engine first.
Obsidian wins decisively. Backlinks, unlinked mentions, graph view, local graph, Markdown composition, attachments, PDFs, canvases, and plugin extensibility make it better for forming and inspecting ideas. It supports “thinking in public to yourself”: outlines, evergreen notes, research pages, meeting notes, project plans, and linked concepts.
Mem0's structures are useful for machines: entities, metadata, facts, temporal signals, and retrieved snippets. That is excellent for personalization and recall, weaker for deliberate synthesis and reflective note-making.
For Hermes, Obsidian can be simple: configure a vault path, let Hermes search/read/write Markdown files, maybe add a convention for inbox notes, promoted memories, project notes, and source-backed research notes. No service has to be alive for the archive to exist.
Mem0 requires operating a memory service or using a hosted API. Self-hosting improves ownership but increases maintenance. Hosted Mem0 reduces ops but increases lock-in and privacy exposure. Mem0 also demands semantic hygiene: what gets written, how conflicts are resolved, when noisy recalled chunks are promoted, and how stale facts are retired.
Obsidian wins. A human can open the vault, inspect every note, edit the source text, review backlinks, and delete bad ideas. This matters because a personal second brain must be trustable under human control.
Mem0 has APIs, dashboards, events, and history, but the actual recall behavior depends on extraction, embeddings, ranking, metadata, and temporal logic. It is observable enough for developers, but less transparent than “open the Markdown note and edit the sentence.”
Obsidian has the stronger default privacy and portability story: local files, open formats, no required account, optional E2E-encrypted Sync, and regular security audits. Its main lock-in risk is ecosystem/plugin dependency, not data format.
Mem0's OSS path is portable and self-hostable, with no vendor lock-in if Brian owns the stack. The hosted path is more convenient but has the usual API/platform risks. Even self-hosted, embeddings and memory extraction create derived data stores that are harder for a human to audit or migrate than Markdown.
Choose Obsidian as the default second brain for Hermes.
Reason: the phrase “second brain” implies more than recall. It means a durable, inspectable, user-owned knowledge workspace. Hermes native memory should stay small and curated; Obsidian should hold larger human-readable knowledge; Mem0 should not become the canonical archive unless Brian wants a machine-first memory system over a human-first one.
Choose Mem0-first only if the primary problem is AI recall quality rather than personal knowledge management: many agent interactions, lots of repeated preferences, cross-agent personalization, semantic lookup over noisy conversation history, or production-like agent memory. Even then, keep Hermes native memory canonical for high-priority rules and keep an export/audit path.
The right answer is a layered combo:
1. Hermes native memory (MEMORY.md / USER.md) — small, curated, high-trust facts and preferences that should influence every session.
2. Obsidian vault — canonical human second brain: notes, projects, research, decisions, evergreen concepts, and source-backed long-form thinking.
3. Mem0 — secondary semantic recall layer: fuzzy retrieval from conversations and agent activity, used to propose candidates for promotion into Hermes memory or Obsidian notes.
Recommended design for Brian:
MEMORY.md and USER.md as the authoritative prompt-injected memory lane.Inbox/ for raw Hermes-captured notes.People/ and Projects/ for durable entities.Research/ for source-backed reports.Decisions/ for canonical “why we chose X” notes.Promoted Memory/ for facts that may deserve native Hermes memory or skills.For Hermes, Obsidian should be the second brain; Mem0 should be the semantic memory sidecar. Obsidian is where Brian's knowledge should live. Mem0 is how Hermes can remember and retrieve fuzzy conversational residue. Hermes native memory remains the curated control plane.