Exomem MCP¶
Self-hosted MCP server (local, no account) — Exomem runs agent memory over the Markdown knowledge base you already own: a plain folder or an Obsidian vault. Claude Code, Codex, and Cursor get durable context with hybrid keyword-plus-vector search, governed review queues, and evidence that stays in plain files. AGPL-3.0, pip install exomem, from Substrate Systems.
Server type: Self-hosted (stdio / local, plus CLI and REST surfaces)
Auth: none — no account, nothing leaves your machine in the lean install
Install: pip install exomem
Tools: search, capture, notes, evidence, audit, and review queues (MCP tools mirror the CLI/REST registry)
Pricing: open source (AGPL-3.0); hosted Exomem is a friends-only private alpha
Category: Memory
Built by: Substrate Systems (substratesystems.io/exomem)
Why This Matters for Operators¶
Every cloud memory service extracts your notes into a vector database you never get back — the memory becomes a derived copy in someone else's infrastructure. Exomem inverts that: the Markdown files are the memory, and the index is a local SQLite sidecar next to them.
The mechanism that matters is governed memory: source material, compiled notes, typed entities, evidence, and supersession history remain plain files you can open, grep, and version. Supersession lives in the file (status: superseded + superseded_by), not in a hidden database — so when a decision gets overturned, the trail is reviewable. Batch embeddings are measured at 256 on 16 GB cards, with methodology published so you can reproduce the numbers on your own vault.
Tools & Capabilities¶
| Area | Purpose |
|---|---|
| Search & capture | Hybrid keyword (SQLite FTS5) + vector (sqlite-vec) retrieval over typed Markdown |
| Notes & entities | Compiled notes, typed entities, and evidence records |
| Audit & review queues | Review queues for captured material and an audit surface |
| Media ingestion | Local OCR, ASR, PDF/Office extraction, and CLIP image indexing |
| CLI / REST | The same operation registry serves CLI, REST, and MCP tools |
Measured on a 50,000-note corpus: 864 ms hybrid find end-to-end (hot cache off), sub-10 ms keyword lanes from FTS5.
Installation¶
pip install exomem
exomem --help
# extras: local embeddings, CLIP, OCR, ASR
Works with Claude Code, Claude Desktop, Codex, Cursor, or any MCP client. GPU is optional — the lean install has no cloud dependency.
Configuration¶
Add the local server to your MCP client pointing at the exomem binary, with your vault folder as the knowledge base. The CLI (kb / exomem) and a personal REST facade expose the same memory for non-MCP workflows.
Business Relevance¶
- Operators with Obsidian vaults get agent memory over files they already own — no migration into a new app.
- Research-heavy teams get review queues so captured material is governed before it becomes durable context.
- Security-conscious operators get a lean install where nothing is uploaded, ever.
- Founders running local AI stacks get memory that is inspectable infrastructure, not hidden assistant state.
Integration with CorpusIQ¶
Exomem complements CorpusIQ's research and governance layers. Market research notes from the research intelligence framework can live in a vault the agent searches with Exomem, while CorpusIQ's canonical facts and metric specs stay the declared source of truth — Exomem's evidence and supersession records make the "declared vs observed" distinction auditable. The file-native design pairs naturally with the docs repo workflow: a memory note, a decision, and its supersession can all travel through git.
Limitations¶
- Brand new — no track record yet; listing appeared August 17, 2026.
- Local-first means no built-in multi-machine sync — the vault is wherever your files are.
- Embedding extras (CLIP, OCR, ASR) add dependencies; the lean install is search-only.
- Hosted option is a friends-only private alpha with no public checkout.
- AGPL-3.0 licensing matters if you embed it in a proprietary product.