Obsidian Mem

中文 (Chinese)

Shared agent memory for DeepSeek Harness, built on plain Markdown vaults.

Obsidian Mem turns your Obsidian-style vaults into a frictionless, human-editable working wiki that every agent session can read from and write to.

OM adds Retrieval-Augmented Generation (RAG) Recall with vector search powered by a Milvus-Lite DB.

Open, edit, and sync your notes with a Markdown setup like Obsidian + Git, then watch your agents learn from shared experiences.

Install

Command line:

# Latest NPM package.
dsh plugin --profile web add dsh-plugin-obsidian-mem

# From source.
dsh plugin --profile web add https://github.com/cybtachyon/dsh-plugin-obsidian-mem

When installing from a local checkout, run pnpm build first: it type-checks the workspace (tsc -b) and then produces the dsh/client.js browser card artifact with esbuild (node scripts/build-client.mjs).

Web UI: click Plugins → Add Plugin → paste dsh-plugin-obsidian-mem.

Or add from the DSH Plugin Marketplace.

Features

• Automatic RAG recall. Lifecycle hooks inject relevant notes at turn start. When the agent encounters highly-relevant files or thinking keywords, that context (capped to a reasonable amount) is also injected. At turn-end, a reminder tells the agent what is worth saving.
• Five agent tools.
  • memory_read: open a memory note from a vault by path.
  • memory_search: hybrid vector + keyword recall across your vaults.
  • memory_write: save a memory note (shared vaults only).
  • memory_audit: report vault size, index state, and over-budget warnings.
  • memory_dream: run the split-compaction pass that merges and prunes notes.
• Works out of the box. Whether you have dozens of vaults, or just one or two AGENTS.md bullets that should be shared, you can configure your own shared vaults. Or just use a default shared vault at $HOME/.dsh/obsidian-mem/!
• Vault permissions. Safeguard your Markdown as a vault as Read-only (agents query but never write) or work together with Adopt and Share (agents read and write).
• Configurable settings. The "Obsidian Mem" card in the Web GUI manages your list of vaults and file preferences. Adjust search and set up automatic Dreaming on a schedule. Of course, you can also ask your agent to set all of these up for you in DSH Creator mode too!
• Global memory file. Vaults support a configurable, editable global memory file. They're added at session start within the global memory budget.

Why another agentic memory project?

Out of the many, many out there (and I looked at as many as I could), none of them:

• let my add as many Markdown / Obsidian Vaults as I wanted
• whenever I wanted
• plus greatly customize the organization
• AND have vector search + maintenance.

I wanted to share project notes with my agents and have maintained results be the cumulative effort of human and automated knowledge organization.

Storage layout

All *.md files are indexed, except: dot directories and dot files are never indexed, and these vault-root paths are reserved for the plugin:

Vault Path	Description
<vault>/.<slug>_history.md	Note history retired by dreaming; reserved suffix.
<vault>/.INTENTS.md	Agent intents file describing phrases that lead to memories.
<vault>/.obsidian-mem/metadata/	Indexer state file and the DREAMS.md dream diary.

Plugin Path	Description
$HOME/.dsh/obsidian-mem/	The default shared vault (created if none are already added).
$HOME/.dsh/obsidian-mem-db/	The memory search Milvus-Lite database (rebuildable on-demand; safe to delete).
$HOME/.dsh/obsidian-mem-db/models/	The memory search ONNX model cache.

Configuration

The settings card is available in the DSH plugin configuration:

Knob	Default	Notes
vaults[]	$HOME/.dsh/obsidian-mem/	Each vault: path, mode (shared | readonly), globalFile (vault-root relative; blank = GLOBAL.md)
vectorDim	384	Full e5-small width; 128..320 are lossy; change = re-index
embeddingModel	Xenova/multilingual-e5-small	ONNX q8, passage: / query: prefixes
modelCacheDir	$HOME/.dsh/obsidian-mem-db/models	transformers.js env.cacheDir
dbDir	$HOME/.dsh/obsidian-mem-db	Milvus-Lite data directory
keywordLlm	off	Optional OpenAI-compatible endpoint that augments heuristic wake-keyword extraction
recallBudget	per plan	Caps injected recall context per turn

Technical Background

The plugin manages its own vector database on-demand:

1. Ensures a uv/uvx toolchain (auto-installing uv when absent)
2. Spawns a local milvus-lite server bound to a localhost port with data under $HOME/.dsh/obsidian-mem-db/

There is no user-facing Python, Milvus install step, or external Milvus.

RAG is a Hybrid search with both dense (HNSW/COSINE) search and sparse BM25 search powered by a local ONNX model. chunk_text is stored without embed-time prefixes and uses enable_analyzer.

Model download and offline use

The plugin downloads Xenova/multilingual-e5-small (ONNX q8, roughly 135 MB) into $HOME/.dsh/obsidian-mem-db/models. For an offline host, pre-seed the cache on a connected machine:

node scripts/prefetch-model.mjs

Development

Build, test, and lint with pnpm; see AGENTS.md for the architecture and the verification commands.

• pnpm test is fully offline: a second integration tier (packages/service/tests/integration/*.live.test.ts) is collected only with RUN_LIVE=1 pnpm test.
• The live tier spawns a real uvx-managed milvus-lite (first run downloads it and can take several minutes).
• Live loads the real Xenova/multilingual-e5-small model from the local cache and drives the full provisioning → indexing → hybrid search → write-through → purge pipeline.
• Per-test timeouts are a generous 180 s to absorb cold caches.
• Everything it writes lands in temporary directories; only the model cache is reused.

Run it with:

RUN_LIVE=1 pnpm test

Reference and Prior-Art

• The excellent Tinqiao Engramory Inspired a majority of the Markdown system design and memory management.
• Zilliz Memsearch Milvus search, reranking.
• bufan528 Memory Vault Recall budgeting, metadata, extraction design.
• The massive amount of DSH Markdown, RAG, and memory plugins, most of which were very interesting but did not provide useful work materia. Still, cool to see the community innovating and trying different things, and I learned a lot about what I did not like or want to use.