BrainLLM

Give Claude a memory that survives the session.

Self-hosted memory for LLMs in TriliumNext Notes, structured by the server and editable by you, served over the Model Context Protocol.

Site · Connect to Claude · How it works · Docs · Privacy

Version Sponsor

What is BrainLLM?

LLMs forget. Every session starts from zero: who you are, what you're working on, what was decided yesterday, what went wrong last week. BrainLLM fixes that.

BrainLLM is an MCP server that gives an LLM a real, persistent memory — stored in TriliumNext Notes, a mature open-source knowledge base you self-host and own. The model opens each session by loading who you are and what's live, writes durable facts the moment they surface, wires knowledge together as a typed graph, and closes each session with a log, a diary entry, and a database backup. The next session picks up exactly where the last one ended.

It's a single Bun/TypeScript service with two dependencies (the MCP SDK and Zod), speaking to Trilium exclusively through its public ETAPI. Your memory lives in your Trilium instance — inspectable, editable, and portable, never locked inside a vendor's black box.

Who it's for

Anyone running an LLM as a long-running collaborator, not a one-shot Q&A tool — across coding, research, personal knowledge management, or day-to-day operational tracking — who wants that collaborator to actually remember: who you are, what's in flight, what was decided, without you re-explaining it every session. If you're comfortable self-hosting one small service and a Trilium instance, BrainLLM turns "the model forgot everything again" into a solved problem. It works with any MCP-capable client (Claude Desktop, Claude Code, or anything else speaking MCP), not just one product.

From experiment to open source

BrainLLM began in February 2026 as a personal experiment: could an LLM operate a real, self-hosted memory reliably enough to be trusted as its own memory — orienting, writing, connecting, and closing sessions without a human doing the filing? Through sustained daily, production use the answer held. The design has settled, the failure modes have been found and fixed, and the project has graduated from experiment to something efficient and stable enough to share — so it's now open source. It still runs the author's own sessions every day; what you're reading is the same code, not a demo. It also still carries the shape of one person's daily use — see Adapting to your environment for exactly what that means and what to change.

The core principle

The model supplies content. The server owns form.

Placement, naming, labels, deduplication, relation bookkeeping, lifecycle aging, archival, date stamping, HTML sanitization, backups — and structure itself — are all deterministic server policy, never delegated to the LLM. The model never chooses a parent note, never sets a label, never checks for duplicates, never stamps a date. That division is what makes the memory reliable: every guarantee is enforced at the tool layer, not requested via prompt.

Design highlights

• Structure is enforced, not requested — every content kind has a canonical structure, served by template() and held on write: a new thread requires its goal, thread/diary/session entries open with an identification line (which LLM, which environment, which session type), threads carry exactly one Resolution (owned by resolve()), duplicate section headings are detected, and Last updated stamps are server-maintained.
• Domains born complete — creating a knowledge domain creates its book and its canonical Sources note (marker legend, stamp, grouped source list, revision table), so every claim has a sourcing home from the first write.
• A visible graph — graph() renders the whole relation graph (or any note's neighborhood) as a Mermaid flowchart, maintained as a native Trilium note.
• Byte-preserving artifacts — text is UTF-8, while binary content uses strict standard base64 in the tool contract and raw bytes on the wire; binary reads return an explicit MIME/base64 envelope instead of replacement characters.
• One-call day orientation — day() serves the previous session, its change log and everything touched since in a single call.
• Two thread shapes — a thread is a dated journal (one child per active day) or a titled collection (one maintained note per item, e.g. an ideas list), chosen once with shape= at creation.
• Timeless notes, dated records — knowledge, information and thread books hold no state or history (only a domain's Current State does); maintain(deep=true) flags dated prose that creeps in. Every note except a log carries an icon, set by kind and backfilled by the sweep.
• Resilient plumbing — every backend call is timeout-bounded with retry on idempotent reads; routine core writes are idempotent or duplicate-guarded, while full-mode raw overwrites remain explicit; content surgery survives the editor's own HTML rewriting; renaming a domain cascades to everything inside it; the maintenance sweep heals drift it finds.
• Multi-agent by default — all write-classified tools serialize behind a process-wide FIFO lock, so two agents (an interactive session and an automated run, say) writing through one hosted instance queue in arrival order instead of racing; reads stay fully parallel. Deletion catch-up in the daily log keeps a note deleted after its day's close from vanishing without a trace — and its window is configurable (deletionCatchupDays in the lifecycle policy), defaulting to Trilium's 7-day retention.
• Ask in prose, split on a seam — consistency(subject="…") finds every note asserting about a fact however it is phrased, no regex guessing; consistency(pattern, staleAfterDays=N) surfaces figures held in exactly one untouched note, the ones that rot silently because nothing disagrees with them, and skips sessions and diaries, which are history rather than claims; diff(since="today") reviews every note a session changed in one call; revise(find="Phase 2", within="tr") edits, inserts or removes a table row named by a few words inside it; read(ids=[…]) batches a multi-note orientation into one round trip — each note's size up front, or a single section= across them; and split(noteId, sections=[…], into="…") is the write half of the oversized-note problem — it lifts whole sections into a new note and leaves a pointer back. Information notes can carry a #mandate marker so a scoped session finds the one note it must obey without reading every note's prose.
• Registers close in one call — revise(find, within="tr", mode="remove", closure={thread, body}) removes a register row and records its closure in that thread's entry for the day, closure={note} moves the row into another register's table shaped to it, and append={table, row} adds a row with no anchor to find; within= also accepts a row as a reader sees it ("Phase 2 | Claude | Open"), and a table row whose cells disagree with its header is flagged at write time and by maintain(deep).
• Records you can prove — sessions, diary entries and dated thread entries are append-only, and now verifiably so: every close() seals its day's records into a digest chained to the previous day's, and maintain(deep) reports any sealed record that was rewritten afterwards. Appends stay legal; edits to history do not go unnoticed.

How it works

At bootstrap, BrainLLM builds a five-area tree in Trilium. Every note the tools create is typed, labeled, dated, and placed by server policy:

BrainLLM  (#brainLlmRoot)
├── 👤 Master       Biography · Goals · Preferences               (the user — maintained singletons)
├── 🤖 LLM          Responsibilities · Protocols · Self-correction · Diary/  (the assistant's self-model + daily diary)
├── 🗂️ Memory       Sessions/ · Threads/                          (daily session logs + multi-session work)
├── 📚 Knowledge    Master/ · Domains/<domain>/{ Sources, info }  (learned facts beyond/contra training)
└── 💡 Insights     Logs/ · Graph · Claims                         (the brain's record of itself)

A session follows an enforced protocol: start() orients (who you are, live threads, what changed since last time) → the model works, writing durable facts as they surface → session() → addendum() → maintain() → remarks() → diary() → close() commits the log, regenerates the daily change log, and triggers a DB backup. The pre-close gate is enforced in code: close() refuses until every step actually ran, in order — narrating "I did the steps" doesn't count, only tool calls do. The gate is durable, written to the session note rather than held in memory, so it survives a restart mid-session and behaves the same on stdio and HTTP. A session that wrote no content gets a light close: nothing to fold or sweep, so close() asks only for session(), remarks() and a short diary().

Orientation is deliberately cheap. start() returns your singletons as section headings — enough for the model to know what the brain holds — and it pulls the one section that matters. Serving the entire self-model on every session meant a one-line question cost the same as a day's work; start(depth="full") is still there for the sessions that genuinely need it.

Knowledge is a typed graph: a closed vocabulary of 16 relations (extends, contradicts, supports, partOf, supersedes, corrects, …), wired by connect() or at creation, traversed by explore() (links / backlinks / neighborhood / shortest path), rendered by graph(), and audited by maintain(deep), which heals duplicate edges and flags what a note application would never think to look for: orphaned notes, stale content, duplicate titles, structural drift inside a note, bodies past the read ceiling, stubs that were labelled and never written, titles carrying a date (which defeats dedup-by-title), threads heavy enough to want consolidating, Sources notes whose verification table was never filled in, and bodies carrying doubly-escaped markup. Most of it is computed from note properties server-side rather than by reading bodies, so the report is complete rather than truncated. Findings can be acknowledged (maintain(ack=[…])) so a flag you have judged correct goes quiet until that note's content actually changes — a warning that reappears every run and is correctly ignored every run trains you to skim the list.

Quick start

Prerequisites

• Bun v1.0+
• A running TriliumNext instance (desktop app or server)

1. Install and build

git clone https://github.com/miisodev/BrainLLM
cd BrainLLM
bun install
bun run build

2. Get an ETAPI token

Or skip this step. Set TRILIUM_PASSWORD instead of TRILIUM_ETAPI_TOKEN and BrainLLM mints a token itself on first start, caching it beside brainllm.json so restarts reuse it rather than creating a new one each time. This is the easiest path for a container deploy, where there's no terminal to run a setup command in.

To do it explicitly instead, either let init mint one:

TRILIUM_BASE_URL=http://localhost:8080 TRILIUM_PASSWORD=your-trilium-password bun run init

It caches the token beside brainllm.json with owner-only permissions and does not print the secret. If you explicitly need a copy, run bun run init -- --show-token; the CLI is idempotent and adopts an existing #brainLlmRoot instead of creating a duplicate tree. Or create one by hand in Trilium: Options → ETAPI → Create token.

3. Configure your MCP client

For Claude Desktop, add to claude_desktop_config.json (see config.example.json for a complete example including remote):

{
  "mcpServers": {
    "BrainLLM": {
      "command": "bun",
      "args": ["run", "/absolute/path/to/BrainLLM/dist/index.js"],
      "env": {
        "TRILIUM_BASE_URL": "http://localhost:8080",
        "TRILIUM_ETAPI_TOKEN": "your-etapi-token-here"
      }
    }
  }
}

For Claude Code, the same server block works in .mcp.json. A schema-valid MCPB manifest (manifest.json) is included for hosts that support custom Bun commands, but BrainLLM still requires Bun and this repository does not yet publish a .mcpb one-click artifact.

4. Bootstrap the brain

Restart your client and ask Claude to run bootstrap — it builds the five-area tree (each area engraved with its purpose), writes brainllm.json, and is active immediately. Or from the CLI:

TRILIUM_BASE_URL=http://localhost:8080 TRILIUM_ETAPI_TOKEN=your-token bun run init

5. Teach the model to operate it

Install the operational skill — the document that teaches Claude to treat BrainLLM as its own mind rather than a filing cabinet:

# Claude Code / Cowork skills directory (macOS/Linux)
cp -r skills/brainllm ~/.claude/skills/brainllm

# Windows PowerShell
New-Item -ItemType Directory -Force "$env:USERPROFILE\.agents\skills" | Out-Null
Copy-Item skills\brainllm "$env:USERPROFILE\.agents\skills\brainllm" -Recurse -Force

Then start a session. The model calls start(), orients, and operates the brain natively from there.

The tool surface

All tool returns are structured JSON. User-input mistakes return informational errors ({error, detail, hint}) the model can read and self-correct from; every read/write/search tool includes a free relations snippet; all writes are retry-safe.

Why the tools are named the way they are

The naming is a readability feature for you, not a style choice. Most MCP clients show you the name of each tool call as it happens. BrainLLM's tools are named as plain-English verbs for what is happening to your memory — remember, recall, revise, resolve, forget, connect, maintain — so that the call log reads as a description of what the model is doing in your brain, without you having to know the schema or open Trilium to find out.

The alternative, which most servers pick, is to name tools after the API underneath: create_note, patch_note, update_attribute. That tells you a row changed. It doesn't tell you the model recorded a decision, corrected something it previously believed, or closed a line of work — and those are the things you actually want to see scroll past.

Read this way, a session is legible at a glance:

Three conventions keep that legible as the surface grows:

• Verbs for actions, nouns for surfaces. remember writes; knowledge reads. If a tool name is a noun, it only ever reads.
• <surface> / <surface>_recall pairs. The bare name reads a surface in full; the _recall suffix skims it. Seeing memory_recall rather than memory tells you the model chose the cheap read.
• Destructive verbs say so. forget, detach, delete_* are never spelled as something softer, so an irreversible call is never disguised as a routine one.

The full mode tools deliberately break this convention — they are named after the raw Trilium primitives (create_note, patch_note, delete_attribute) precisely so that a raw, guard-free operation is visibly different in the log from a brain-aware one.

Tool permissions

Every tool declares whether it reads, writes, or destroys, so your client can group them and you can grant permission per group instead of per call. Every tool also carries a human-readable title ("Edit a note", "Search the brain") for permission prompts and directory listings. The current full-mode split is 34 read-only · 31 write · 13 destructive (core mode is 23 read-only · 17 write · 5 destructive); revise, label and connect count as destructive because they can overwrite or remove what was there, which lets reads run unattended while anything that touches your brain still asks.

The classification lives in one reviewable table (src/annotations.ts) rather than scattered across registrations, and its default is deliberately unsafe-side: a tool missing from the table is treated as a write, never a read, and says so at startup. Three are worth knowing because they look like reads and aren't — start() creates today's diary and session stubs, session() runs the maintenance sweep, and graph() writes the rendered graph note.

Core — universal verbs (35)

Core — surface reads (10, dual-mode)

master/master_recall · llm/llm_recall · memory/memory_recall · knowledge/knowledge_recall · insights/insights_recall — each surface read in full, or skimmed as compact stubs.

Full mode (BRAINLLM_MODE=full, +33)

The raw ETAPI, one tool per Trilium primitive — notes, attributes, branches, revisions, attachments, calendar, system. Brain-agnostic and guard-free: an edge-case fallback for what core can't express (precise raw queries, code/canvas/mermaid notes, branch surgery, revision recovery).

The complete operational reference is skills/brainllm/SKILL.md; per-topic deep dives live in skills/brainllm/references/. Beyond that, the architecture reference explains how BrainLLM is built and why — every module, every tool's behaviour, the sweep internals, the idempotency matrix — and is kept current with each release rather than published in this repo; see Support the project for how to get a copy.

Deployment

Transport modes

The HTTP connector serves a streamable-HTTP MCP endpoint at /mcp (one session per mcp-session-id, DELETE terminates, CORS-enabled with mcp-session-id exposed for browser clients), a legacy SSE endpoint at /sse (with POST /messages) for clients that predate streamable HTTP — older Cursor builds, Continue, and Python-SDK clients — and GET /health. The server root serves a plain page for ordinary browser GET requests; MCP-shaped requests sent to / are also accepted as a compatibility alias, because some hosted Claude connector versions normalize the post-token MCP path to the origin root. The server itself speaks plain HTTP — TLS is expected to terminate in front of it (Railway's edge does this automatically for you; on a bare VPS or your own Docker host, put it behind a reverse proxy such as Caddy, nginx, or a Cloudflare Tunnel). HTTP mode now fails closed unless MCP_AUTH_TOKEN or BRAINLLM_OWNER_PASSWORD is configured; BRAINLLM_ALLOW_UNAUTHENTICATED_HTTP=true is an explicit escape hatch for an isolated trusted network only. CORS defaults to *, so a reachable deployment must have a real authentication gate. Idle sessions are evicted after 1 hour; MCP request bodies are capped at 50 MiB by both the server and SDK transport. Both transports sit behind the same authentication gate.

Docker / Railway

The included Dockerfile builds and runs the HTTP connector (two-stage, digest-pinned oven/bun, drops privileges via entrypoint.sh). On Railway: PORT is auto-injected — set MCP_AUTH_TOKEN, TRILIUM_BASE_URL, TRILIUM_ETAPI_TOKEN, and BRAINLLM_TZ as service variables, deploy, and point your client at https://<app>.up.railway.app/mcp with the bearer token. Deploying elsewhere (a VPS, your own Docker host) works the same way — the container only needs those same env vars and a TLS-terminating proxy in front of it, per the transport note above.

Branding

In HTTP mode BrainLLM serves its own icons — /icon.png, /icon.svg, sized variants, and /favicon.ico — and advertises them in the MCP handshake together with its title, a one-line description and its website, so clients show the BrainLLM mark in your connector list and beside its tool calls. The handshake also carries short server instructions describing the five areas and the session shape. Everything brand-shaped (name, tagline, description, the mark's geometry) is defined once in src/brand.ts, and a unit test holds package.json, manifest.json and server.json to the same description.

Both halves of that are load-bearing. The spec tells clients to verify icon URIs share the server's origin, so icons hosted anywhere else are silently dropped. And a host that serves no favicon makes clients fall back to the registrable domain's — a BrainLLM at brain.yourdomain.com would show yourdomain.com's site logo, which is worse than nothing because it looks intentional.

To use your own artwork, replace public/BrainLLM.svg and run bun run icons — it rasterises the vector at each size the server serves. Commit the results; the Docker image copies public/ and has no image toolchain at runtime.

Connecting remote clients (OAuth)

Claude Code and mcp-remote send MCP_AUTH_TOKEN as a header and need nothing further. The hosted Claude surfaces — claude.ai, Claude mobile, Cowork — cannot. Their custom-connector UI offers OAuth or nothing; there is no field for a bearer token. So BrainLLM ships its own OAuth 2.1 authorization server. It serves both registration mechanisms, so every OAuth-capable client works:

• Client ID Metadata Documents (CIMD) — the registration-free mechanism adopted in the MCP 2025-11-25 specification; this is what Claude selects.
• Dynamic Client Registration (RFC 7591) — for clients that never grew CIMD support. opencode (MCP TS SDK ≤1.29) is the forcing case: without a registration_endpoint it refuses to authenticate at all.

Set one variable to turn it on:

BRAINLLM_OWNER_PASSWORD=<a strong password you choose>

That password is what you type on the consent screen. Without it the OAuth endpoints are disabled entirely and nothing else changes.

Then in Claude: Settings → Connectors → Add custom connector, enter https://<your-host>/mcp, and leave the OAuth client fields blank — CIMD needs no client ID or secret. Claude discovers everything from the 401, you approve once on a consent screen, and it stays connected.

What that turns on:

The consent screen names the requesting client and the address it will return to, and warns when that address is your own computer (a loopback redirect, as Claude Code uses). MCP traffic is rate-limited per bearer token, with a looser per-IP bound, so clients that share an egress range (all of Claude's hosted traffic does) never exhaust each other's budget.

Both credentials work simultaneously — a static token from Claude Code and an OAuth token from claude.ai, against the same brain. Access tokens are signed JWTs bound to your server's resource URI, so a token minted for a different MCP server is rejected; the signing secret and refresh tokens persist beside brainllm.json, which on a container deploy means putting BRAINLLM_CONFIG on a volume — otherwise every redeploy invalidates every token.

For config persistence across redeploys, mount a volume on the BrainLLM service (not the Trilium service) and set BRAINLLM_CONFIG to a file path inside it:

railway volume -p <project-id> -s <brainllm-service-id> -e production add --mount-path /vol
railway variables set BRAINLLM_CONFIG=/vol/brainllm.json

Without a volume, leave BRAINLLM_CONFIG unset — auto-discovery re-finds the brain from Trilium's #brainLlmRoot marker on each cold start (~1 s).

Configuration reference

Lifecycle timings live in brainllm.json and are yours to tune:

"policy": { "dormantAfterDays": 21, "archiveDormantAfterDays": 45, "staleAfterDays": 7 }

Threads age active → dormant → archived-in-place on that timeline; nothing is ever deleted by aging, and recover() / withdraw() bring anything back.

Adapting to your environment

A handful of values in this repo reflect the author's own machine. None are secrets, but if you clone or fork, review these and swap in your own:

Architecture

index.ts ─→ tools.ts ─┬→ trilium.ts     ETAPI client: bounded/retrying I/O, graph traversal
                      ├→ router.ts      placement policy: kind → parent, label plans, domains
                      ├→ templates.ts   canonical structures per kind + the template-tool rules
                      ├→ normalize.ts   titles, slugs, markdown→HTML, sanitization, surgery
                      ├→ lifecycle.ts   protection, sweep, aging, the start digest
                      ├→ journal.ts     daily Insights log generation (regenerate-in-place)
                      ├→ time.ts        timezone-correct now/today (BRAINLLM_TZ)
                       ├→ rate-limit.ts  bounded public HTTP budgets
                       ├→ brand.ts       name, description, mark, icons, server instructions
                       ├→ cap.ts         the 140k-character result cap
                       ├→ diffing.ts     block and table-cell diffs
                       ├→ network-security.ts CIMD public-network and redirect checks
                      ├→ bootstrap.ts   five-area tree builder
                      └→ tools-*.ts     per-area surface reads · full-mode raw ETAPI

Key properties: no tool result exceeds 140,000 characters (a larger one is cut with a note naming the narrower read); every write is sanitized for Trilium/CKEditor 5 compatibility (mutations reported back as sanitized[]); a revision snapshot precedes routine content mutation; every request carries the user's local time so Trilium stamps dates in the right day; core writes are idempotent or duplicate-guarded, while raw full-mode overwrites are explicit.

Development

bun run dev    # hot-reload dev server
bun run build  # bundle to dist/index.js
bun run test:unit   # all unit tests (Bun discovery; works on Windows and CI)
bun run test:http   # local auth/header smoke (no Trilium required)
BRAINLLM_LIVE_TOKEN=... bun run test:live   # read-only smoke against a deployed server, over real MCP
bun run test        # integration tests (requires a live Trilium instance)
bun run validate:manifest # MCPB schema/icon validation
bun run init   # CLI bootstrap

Repo layout: runtime source in src/, the operational skill in skills/brainllm/, developer scripts in scripts/, schema-valid MCPB manifest in manifest.json, container build in Dockerfile.

Contributing

Contributions are welcome — bug reports, fixes, docs, and ideas alike.

What we value

BrainLLM started as one person's daily-driver tool, so the code and docs still carry that person's assumptions in a few places (see Adapting to your environment). The contributions that matter most are the ones that generalize it beyond that:

How to contribute

1. Open an issue first for anything non-trivial — bugs with reproduction steps, or proposals with the use case spelled out.
2. Fork and branch, keep changes focused, and match the existing code style.
3. bun run build must pass clean and unit tests must stay green (bun run test:unit — CI runs the same checks on every push and PR); add tests for new normalize/lifecycle logic.
4. Use conventional commits (fix:, feat:, docs:, refactor:).
5. Update the docs that your change touches — README.md, the skill package under skills/brainllm/, and config.example.json/.env.example where relevant. The skill is part of the product: a tool change without its skill update is half a change.

Not sure where to start? Issues labeled good first issue, doc gaps, and the troubleshooting reference are all friendly entry points.

Security

Please do not report suspected vulnerabilities in a public issue. Follow the private reporting process in SECURITY.md; never include ETAPI tokens, OAuth credentials, or private brain content in a report.

Privacy and support

BrainLLM collects nothing: no analytics, no telemetry, no account. Your notes stay in your Trilium, and sign-in records stay on your own server. The full privacy policy says what is stored where. For help, open a GitHub issue or email miisodev@gmail.com.

Support the project

BrainLLM is built and maintained by one person. If it's useful to you, sponsorship directly funds its continued development:

Sponsors can request a copy of the Blueprint — the complete internal architecture reference generated from full source analysis (every module, every tool's exact behavior and signatures, the sweep and idempotency internals) that isn't published in this repo. Message via GitHub Sponsors to ask for one.

Starring the repo, reporting bugs, and spreading the word help too.

License

MIT © Kevin Miiso Novo

Credits

• TriliumNext Notes — the open-source, self-hosted knowledge base that powers this server's backend. BrainLLM would not exist without the TriliumNext team's work.
• Model Context Protocol — the open standard that lets one memory serve every MCP-capable client.