Memories

Memories are notes you keep in Loca for the agent to draw on. Unlike an attachment, which belongs to a single message, a memory lives on your Core and stays available across every session, so a reference you add once is there for every future conversation. They are how you give the agent lasting context about your work, your projects, or yourself.

Knowledge and skills

A memory is one of two kinds, and the difference is what the agent is meant to do with it. Knowledge is reference data: facts, notes, a spec, a list the agent reads to answer accurately but never treats as instructions. A skill is the opposite, a set of procedural instructions the agent follows when the moment fits, like a checklist for a recurring task. In short, knowledge is what the agent knows, and a skill is what the agent does.

A skill can carry more than a single page. Where a piece of knowledge is one note, a skill can bundle reference files alongside its instructions, and the agent pulls those in only when it actually works through the skill. That keeps a detailed procedure from crowding the conversation until it is needed.

Enabling a memory

Every memory has an on/off toggle, so you can keep one written down without it being in play. Turn a memory off and it stays saved but is set aside: Loca will not surface it and the agent will not draw on it until you turn it back on. This is the clean way to park a note you are not using right now without deleting it.

Relevant memories

You do not have to remember which memory matters for a given question. As you write a message, Loca finds the memories most related to it, lets the agent see their summaries, and lets it read the full text on demand. The memory surfaces itself when it is relevant, and stays out of the way when it is not.

The conversation shows which memories were surfaced, so a reply that leans on one tells you which it used. When nothing matches, Loca can either stay quiet or show that it looked and found nothing, whichever you prefer, so silence is never ambiguous.

Skills surface the same way, by relevance, and you can also reach for one deliberately whenever you want it. If you would rather a skill only ever run when you ask for it, turn off skill auto-relevancy in Settings; the manual trigger stays available either way.

Referencing a memory

When you know exactly which memory you want, you do not have to wait for it to be ranked in. Type @ and the memory’s name in the input bar to drop its full content straight into your message; @ is the symbol for memories. That is the difference between the two paths: retrieval is Loca finding a memory for you, while a mention is you naming one outright.

Settings

Turn relevant memories on or off, and cap how many Loca offers at once, under Memories in Settings. The cap is the main dial worth knowing: a higher one surfaces more memories per message at the cost of more of the context window, while a lower one keeps replies lean. The same group holds the toggle for skill auto-relevancy, so you can let skills surface on their own or keep them to manual triggers only.

You do not have to write every memory yourself. An extension can ship its own knowledge and skills, installed read-only alongside yours, so adding an extension can hand the agent a body of reference and a set of procedures in one step.

How It Works

Relevant memories use the same idea as the tool shortlist. Each memory carries a short summary, which Loca turns into a list of numbers that captures its meaning ahead of time. When you send a message, Loca does the same to your message and compares it against every memory’s summary, keeping the closest matches that clear a minimum similarity. Those few are handed to the agent, summary first with the full text a step away, up to the cap you set. A skill is ranked the same way when auto-relevancy is on, which is why a relevant procedure can surface on its own, and why naming one with @ or # is you skipping the ranking and pointing at it directly.

For shorter pieces of reusable text you trigger by name rather than memories the agent retrieves, see Snippets.