Engram Chat

Memory

• Memory lives in Intelligence/Memory.md by default. It's a normal note you can read and edit.
• Run /memory to pull key facts from the current chat. You review them before they're saved.
• Automatic saving after each reply is off by default (it costs one extra model call per reply). You can turn it on under Advanced settings → Auto-extract memories.
• The AI can add (save_memory), change (update_memory) or delete (delete_memory) a single entry. Only that entry's line is touched; everything else in the file, including your own headings and text, stays exactly as it is.
• The memory file is never trimmed. The Max memory tokens setting (4,000 by default) only limits how much memory is sent to the model with each message. When it is exceeded, the oldest entries are left out of the prompt, not out of the file.
• Safety: if the memory file goes missing (for example while a sync is still running), Engram does not create a blank one. New facts are queued and added when the file is back. Before every change a local backup is saved in the plugin folder (the last 20 are kept). Use the command Restore memory from backup, or the button in settings, to recover.
• Changing the memory path in settings moves your existing file to the new path after you confirm.

Semantic search

• Embedding providers: Ollama, OpenAI, or any OpenAI-compatible /v1/embeddings endpoint (for example, llama.cpp).
• Click Index Vault in settings to build the index, with live progress.
• Notes used as context show up as clickable chips under your message.
• Changing the embedding model, provider, or URL clears the old index so stale vectors are never mixed in.

Providers

• Local: Ollama, LM Studio, llama.cpp
• Cloud: OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Groq, Mistral, xAI (Grok)
• Aggregator: OpenRouter (one key, many models)
• Custom: any OpenAI-compatible API
• Replies stream in. Each provider has its own API key; a key is only ever sent to the provider it was entered for. Claude is called directly using Anthropic's browser-access header. The reply length limit is set in Advanced settings (Max reply length).
• Connection tests only list models; they never send a billable chat request.
• Tool calling can be native, prompt-based (for models without tool support), or turned off.

Safety

• Scope: an allowlist or denylist of folders, plus exclude patterns (Private/**, *.secret.md by default). Excluded files are never indexed, searched, or sent to the model.
• Permissions: one click on the chat footer badge switches between:
  • 🔍 Read: search and read only
  • ✏️ Append: can only add text to the end of notes
  • ⚠️ Full edit: can create, edit, overwrite, move, rename, and delete
• Protected files: the AI's tools can never change, move, rename, copy or delete the memory file, the chat history folder, or any folder that contains them. This holds even when memory is switched off.
• Confirmation: destructive actions open a confirmation dialog first.
• Undo panel: review and revert changes the AI made.

Interface

• Type / for commands and @ to insert a link to a note ([[Note]]) into your message. The AI can then open it with its read tools.
• A token budget bar shows how much of the model's context window is in use.
• Personas let you switch between saved system prompts.

ollama pull nomic-embed-text

Ollama runs at http://localhost:11434 by default.

Download a GGUF embedding model, then start the server in embedding mode:

.\llama-server.exe -m "C:\path\to\model.gguf" --port 8080 --embedding

Check that it's running at http://localhost:8080/v1/models.

Use text-embedding-3-small with your OpenAI API key.

• Make sure indexing finished (check the file count in settings).
• Your embedding server must be running while you chat, not just while indexing.
• Look for note chips under your message. If they're there, the notes were sent to the model.

Check the Base URL and that the embedding server is running. Engram tests the connection before indexing and shows an error if it can't connect.

Run .\llama-server.exe (with .\ and .exe) from the folder that contains it, or add that folder to your PATH.

The model returned an empty reply. Try another model or a simpler question. For local models, check that the model is fully loaded and has enough VRAM.

The index is saved between restarts. If it's gone, the embedding model, provider, or URL probably changed, which clears the index on purpose. Re-index once.