Infimium

The Private Context Layer & Super Brain for Your Codebase. Give AI agents persistent memory, deep dependency graphs, and instant code context -- 100% local, zero token bloat.

Demo

Related MCP server: LocalNest MCP

Why

Large repositories make agents read too much code or miss the right symbol. Infimium retrieves compact, relevant context before the agent starts editing.

200,000 lines of code
Agent reads everything -> context blown + expensive
grep "price calculation" -> misses calcPropertyValue()

tool: semantic_code_search
query: "price calculation logic"

-> services/property/calc.ts:142 · calcPropertyValue()
-> callers: getListingPrice(), estimatePropertyTax()

Quick Start

Requires Node.js 22.5+. From your project folder:

cd /path/to/your/project
npx infimium@latest setup

Run setup from the repository you want to index, not from your home directory (~). Infimium stops broad roots automatically so it cannot scan unrelated files.

That creates global config, starts Ollama if it is installed, pulls nomic-embed-text, indexes the current project or workspace, runs doctor, and opens Playground.

The published CLI keeps its executable entrypoint, so MCP clients can launch it directly through the configuration below.

If Ollama is not installed yet:

npx infimium@latest setup --install-deps

infimium setup creates one global config at ~/.infimium/.env. You do not need a .env in every project. Code, docs, memory, graphs, and vectors are stored locally under ~/.infimium/.

Web search is optional. Add a Tinyfish key only when you need it:

SEARCH_PROVIDER=tinyfish
SEARCH_API_KEY=your_key

Full infimium plan generation also needs a local text model:

ollama pull llama3.1

infimium plan --dry-run "your task" works without this model and shows the retrieved code context first.

Connect Your Agent

Cursor, Windsurf, Claude Desktop, and other MCP clients:

{
  "mcpServers": {
    "infimium": {
      "command": "npx",
      "args": ["-y", "infimium", "serve"]
    }
  }
}

Restart the client, then use:

Use Infimium hello_infimium.
Use Infimium get_context before starting.
Use Infimium semantic_code_search to explain this repository.

Infimium normally uses the MCP process working directory. If your client starts it elsewhere, pass project_path once; Infimium remembers the active project and auto-indexes it.

Tools

CLI

Use npx infimium ... if you did not install the package globally.

Project Memory

Infimium keeps memory bounded across long sessions:

• Scratchpad: recent events for the active task.

• Archive: compact summaries of completed tasks.

• Ledger: durable decisions, rules, quirks, and unresolved blockers.

Record meaningful progress while working:

infimium remember "Added rate-limit middleware" --type progress --task "Rate limiting"
infimium remember "Use Redis-backed counters in production" --type decision

When the task is complete:

infimium memory complete

Infimium uses the local llama3.1 model when available and falls back to deterministic compaction when it is not. Raw compacted events remain stored locally for seven days before pruning. get_context never calls an LLM or network service.

From a source checkout, build once and run the local playground with:

npm run build
npm run playground

Local Architecture

• Ollama creates embeddings on your machine.

• Embedded SQLite stores vectors, index metadata, project memory, and graph edges. No ChromaDB or Docker service is required.

• Documents use recursive boundary-aware chunks instead of blind fixed slices.

• JavaScript, TypeScript, Python, and Dart parsers are bundled.

• Go, Rust, and Java Tree-sitter WASM grammars download on first use and cache in ~/.infimium/grammars/.

• .gitignore, .infimiumignore, and framework defaults exclude dependencies, build output, Flutter artifacts, caches, and binaries before indexing.

• semantic_code_search returns signatures; expand_symbol provides full code on demand.

• Project memory uses session-scoped scratchpads, compact milestone archives, and a versioned semantic ledger.

• get_context emits static anchors, dynamic repository state, and active execution as separate YAML zones.

Multiple Projects

Run the normal index command from a folder containing related projects:

infimium index

Infimium detects immediate project roots from files such as pubspec.yaml, package.json, Cargo.toml, and go.mod. It shows the detected roles and dependencies, asks once, then creates infimium.workspace.json, indexes every project, and opens Playground.

For unattended setup:

infimium index --yes --no-playground

Use --no-workspace to index only the current project. Workspace projects keep separate memory and Git state while get_context includes balanced summaries and graph relationships from related projects.

Infimium - Playground

Infimium drops the initial payload cost from approximately 1,460 tokens to 8 tokens per symbol. Semantic search returns the AST signature first; the agent requests the full implementation only when it needs it with expand_symbol.

Full implementation   ~1,460 tokens
AST skeleton               ~8 tokens
Initial payload reduction  ~99.5%

These are Playground reference values, not a claim that every function has the same size. Inspect your own indexed repository and compare AST-first retrieval with full-text retrieval locally:

infimium playground

Open Token Economics to see the estimated token difference across your actual indexed symbols.

Privacy

Code, docs, embeddings, memory, graph data, prompts, queries, file paths, and repo names remain local.

Infimium sends privacy-safe anonymous lifecycle telemetry so we can understand setup success:

• init_started, init_completed

• doctor_run, doctor_passed

• index_started, index_completed, setup_completed

• serve_started, first_tool_call, playground_opened

Telemetry includes an anonymous install ID, Infimium version, OS, Node major version, timestamp, and event name. It never includes code, file paths, repo names, prompts, search queries, memory notes, API keys, or user identity.

Disable it anytime:

infimium telemetry off

or set:

INFIMIUM_TELEMETRY=false

Troubleshooting

FAQ & Common Confusions

Where is layer.md? When you run infimium get-context, it intentionally prints the context directly to your terminal (stdout) so AI agents can read it instantly. It doesn't create a layer.md file in your workspace to avoid clutter. If you want to manually save it to a file, use terminal redirection:

infimium get-context > layer.md

Why does the Playground UI say "Awaiting first agent interaction..."? The CURRENT TASK tracker at the top of the Playground UI is designed to mirror exactly what your AI agent sees. If an agent hasn't queried the context yet (via the get-context tool), it waits. To force it to update, manually run infimium get-context.

How do I format infimium remember? The infimium remember command requires a message and a --type flag (valid types: note, progress, decision, blocker, index, plan). If you also want it to update the active task in the Playground, include the --task flag:

infimium remember "Added rate limiting" --type progress --task "Security Features"

Database is locked

If you see Failed to start Infimium: Database is locked, it means another instance of Infimium is actively holding a lock on the SQLite memory database. This usually happens if you try to run infimium index manually in one terminal while infimium playground or infimium watch is still running in another. Simply stop the running process (Ctrl+C) before running manual commands.

General Setup Issues

Run:

infimium doctor

Every failed check prints one copy-paste fix. If setup still fails, give this prompt to your coding agent:

Set up Infimium in this repository. Install/start Ollama, pull nomic-embed-text,
run npx infimium init, run npx infimium index, and make all six
npx infimium doctor checks pass. Do not commit secrets.

Contributing

See CONTRIBUTING.md. Adding a language starts with a parser fixture and extraction test.

Self-hosting is free forever under the MIT license.