Trove

Trove

Context that keeps getting better.
Open-source MCP memory server: one shared, sourced memory for Claude Code, Codex, Cursor and any MCP client.

Website · Quickstart · Agent usage · MCP tools · OAuth connector · llms.txt

Trove is an open-source MCP server that gives your AI agents long-term memory in a shared knowledge graph. Use it hosted at mytrove.in or self-host it. Claude, Codex, Cursor and any other MCP-capable agent can write what they learn into it and recall it in later sessions — with every fact traceable back to the source text that justifies it. Your notes stay inspectable, contradictions are superseded instead of overwritten, and the whole graph is browsable in a built-in dashboard.

Connect your agent

Point any MCP-capable agent at Trove and it reads and writes your graph across sessions. Create an API key on the dashboard at mytrove.in, then:

# Claude Code (hosted)
claude mcp add trove --transport http https://mytrove.in/mcp \
  --header "Authorization: Bearer <your trove_ key>"

Running your own instance? Use http://localhost:8787/mcp (no key needed when auth is disabled). For the claude.ai / Claude Desktop Connectors panel, add https://mytrove.in/mcp and sign in with Clerk — see docs/oauth.md.

Your agent gets a small verb-per-job toolset: remember, recall, grep, read, connect, forget (plus ingest and curation tools for write-scoped keys). Nothing else to pass — your key already points at your own graph, so all your agents share it.

How agents should use it — not as an end-of-day diary, but as working memory: load before re-deriving, capture mid-session, many small linked atoms. See docs/agent-usage.md. The same doctrine is also on the MCP server (instructions, resource trove://doctrine, tool descriptions).

Companion skills (optional, Claude Code). The MCP tools work in any client on their own. Install the skills straight from this repo so Claude uses them with the right discipline — recall before re-deriving, cite evidence, supersede instead of delete:

npx skills add anunay999/trove -g

Every skill is also served by the API, so any agent can be told to read one and follow it: the index is at mytrove.in/skills.md (and /llms.txt), each file at /skills/<name>.md. The same procedures are exposed as MCP prompts, e.g. /mcp__trove__trove-curate in Claude Code cleans up your graph from inside the session.

Full setup — keys, scopes, stdio, importing a vault — is in docs/quickstart.md.

What you get

• Durable agent memory — agents save facts, decisions, and notes through MCP tools (remember, ingest) and retrieve them later with recall, which packs the most relevant memories into a token budget you set.
• Time-travel, not overwrites — beliefs carry validity intervals. When something changes, the old edge is invalidated, never deleted, so you can ask what the graph believed at any point in time.
• Evidence, by construction — recalled atoms either resolve to cited source text or are explicitly marked AGENT INFERENCE; the linter flags evidence-free nodes (missing_evidence) for review.
• A dashboard — memory KPIs, write-cadence heatmap, lint health, and an interactive force-directed graph explorer with a full-document reader.
• Your notes, imported — point the importer at an Obsidian vault and it becomes the seed of the graph. Append-heavy files like log.md are split into per-entry episodes and deduped, so re-imports only store what's new.

How it works

Trove is an evidence graph: sourced facts trace back to cited text, while evidence-free conclusions stay visibly marked as agent inference.

   raw source            addressable spans        distilled beliefs
  ┌───────────┐  ingest  ┌───────────────┐  remember  ┌──────────┐
  │ transcript│ ───────► │  text units   │ ─────────► │  nodes   │
  │ page/file │          │ (cited spans) │  (cites)   │ + edges  │
  └───────────┘          └───────────────┘            └──────────┘
                                                        ▲   recall / grep / read
                                              agents read back here

Writing. An agent stores raw material with ingest, which splits it into addressable text units. It then distills the few facts worth keeping with remember — small semantic nodes, each citing the text unit that justifies it — and links them with connect. remember is the single write door: it revises an existing node on an exact title/slug match, otherwise creates one, so agents never juggle create-vs-update.

Reading. Three verbs, each for a different question. recall builds a token-budgeted context pack — hybrid lexical+semantic search seeds a one-hop graph expansion, candidates are ranked by relevance plus ACT-R-style activation (recency and frequency), and a greedy packer fills the budget you set. grep is exact/regex search over nodes and raw sources, for when you know the string (a port, a slug, an error). read pulls one node or source by id or slug.

Changing its mind. Facts and relationships keep recorded history instead of being overwritten. Each node revision snapshots its title, summary, and content; read({ asOf }) returns the newest snapshot recorded by that time, or null if the fact did not exist yet. Relationship supersession remains non-destructive: connect with supersedesEdgeId (or forget) expires the old edge, while neighborhood uses asOf for recorded-time edge visibility and validFrom/validUntil for world-time filtering. recall always answers from present belief and rejects asOf. Evidence and annotations on a historical node read currently reflect their present state.

Staying fresh. A background worker inside the server drains a job queue every 30 seconds to embed new writes, refresh projections, and run lint. There's nothing to schedule.

Tiered tools. The MCP surface is filtered by your credential's scope. Read-only keys see only recall, grep, and read; write keys add remember, connect, forget, ingest, and the curation tools; admin keys also see the operator tools (jobs, lint, export_obsidian). Visibility is convenience — every call is still scope-checked server-side.

Get started

Use the hosted service at mytrove.in — sign in, create an API key on the dashboard, and point your agent at https://mytrove.in/mcp. Or run it yourself in about five minutes.

→ docs/quickstart.md — local setup, API keys and scopes, connecting agents (Claude Code, stdio, and the claude.ai OAuth connector), importing an Obsidian vault, and enabling semantic search.

# the short version, locally
git clone https://github.com/anunay999/trove.git && cd trove
npm install                 # server + web dashboard
cp .env.example .env
npm run setup               # Postgres + schema + migrations
npm run web:build && npm start
# → dashboard at http://localhost:8787

Deploy

Trove ships a production Dockerfile (API + dashboard in one image) and a railway.json. The reference deployment is Railway + Supabase Postgres — see docs/deployment.md for the walkthrough, including custom domains.

Learn more

• docs/quickstart.md — setup, keys, connecting agents, importing notes
• docs/oauth.md — the claude.ai OAuth connector (Clerk-backed)
• docs/development.md — architecture, data model, test suites, and design rationale (start here if you're hacking on Trove)
• docs/mcp.md — full MCP tool and resource reference
• docs/agent-api.md — HTTP API contract
• docs/cli.md — the reference CLI
• docs/memory-db-design.md — the research behind the design

Contributing & license

Pull requests are welcome — see CONTRIBUTING.md. CI runs typecheck, the store test suites, and the web build, and main requires a reviewed PR. Please don't include secrets or real connection strings in a diff. Security issues: see SECURITY.md.

Licensed under the GNU AGPL-3.0. You're free to use, modify, and self-host Trove; if you run a modified version as a network service, you must make your source available under the same license.