-
Notifications
You must be signed in to change notification settings - Fork 0
FAQ
No. Recollectium Core is local-first. Data lives in a local SQLite database unless you deliberately place that database somewhere else.
No. Use host-level protections if required, such as OS account permissions, encrypted home directories, encrypted volumes, full-disk encryption, and protected backups.
There is no dedicated backup command in v1. For a manual backup, stop Recollectium first, then copy the data directory or the SQLite database file to a safe location.
If the database is corrupted, restore that copy and start Recollectium again. If you restore an older backup, Recollectium migrates it forward automatically when it opens the database. If the database is newer than the installed build, you need a newer Recollectium release or a backup from before the newer schema was written.
No. The v1 API, MCP, and WebUI services are unauthenticated and localhost-first. Do not expose them publicly.
No. The WebUI is a local browser control plane and uses the same unauthenticated local service contract as Core. Keep it on 127.0.0.1 unless you intentionally add private-network controls.
The default built-in local profile uses BAAI/bge-base-en-v1.5 through FastEmbed. It has 768 dimensions, profile builtin-fastembed-bge-base-en-v1-5-v1, max tokens 512, chunk tokens 384, and overlap 64.
The legacy supported alternative is jinaai/jina-embeddings-v2-small-en. It has 512 dimensions, profile builtin-fastembed-jina-v2-small-en-v1, max tokens 8192, chunk tokens 6144, and overlap 512. Switching embedding model or profile can require existing memories to be refreshed. Recollectium refreshes stale embeddings inline when triggered, and you can force a refresh with embedding-refresh or clear job audit records with embedding-jobs-clear --yes.
Usually no. Search by scope first: user or workspace.
Recollectium uses embedding-based semantic search, so it returns memories that are closest in meaning to the query. In many cases, type is already implied by the words in the query. For example, searching user memory for "preferences about wording" will naturally pull preference-like memories even without --type preference.
If the result set is too broad, narrow by type. Type filters are useful when the agent knows it only wants a specific bucket, such as decision, configuration, or bug_finding.
recollectium dev serve runs the local HTTP API in the foreground. It is useful for development, debugging, and seeing logs directly in the terminal.
recollectium service start api starts a managed background API service, writes PID and discovery metadata, and is the recommended path for adapters and plugins.
The WebUI is the browser control plane for local operators. The API is for HTTP clients and scripts, and MCP is for agent clients that support Model Context Protocol. All three talk to the same Core.
Yes, but treat it as an advanced setup. The API and MCP services are unauthenticated in v1. Use private networking and external access controls, such as Tailscale, WireGuard, SSH tunnels, or firewall allowlists.
The OpenCode adapter remains planned for a future release. The Core service, CLI, API, MCP tools, workspace UID rules, and plugin contract are already available as the foundation for adapter work, but this repository does not ship an OpenCode plugin implementation yet.
FAQ
- Is Recollectium cloud-hosted?
- Does Recollectium encrypt the database?
- Does the API have authentication?
- Does the WebUI have authentication?
- Which embedding model does it use?
- Should agents search by type first?
- What is the difference between
dev serveandservice start api? - Can I run Core on one machine and an agent on another?
- Is the OpenCode plugin available?