Auto-recall

On every prompt, Neuronz.ai searches your stored facts, past work and knowledge docs for what you just asked. It doesn't paste the results into the prompt. It tells your agent how many records match and that it should read them with recall before answering or acting.

This page uses harness for the coding agent you use (Claude Code, the Claude desktop app, Oh-My-Pi). Auto-recall runs wherever the Neuronz.ai plugin is installed: Claude Code, including the desktop app's Code tab, and Oh-My-Pi.

What arrives with each prompt

• Matching records: a count of the facts, past work and knowledge that match, with the call that reads them.
• Rules: the names of the rules that may apply, with the call that fetches their text. The text itself is not sent.
• Your voice: a short note saying how you want the agent to write, which the agent reads once and again only when you change it.
• Persona traits: your own traits come with the session-start briefing and are repeated every 20 prompts and after the agent compacts its context. When a prompt names a colleague, the agent is told who they are and how many traits are stored, and reads them with get_persona.
• Related topics: the titles of a few topics related to your prompt. Topics are not part of recall; the agent opens one with get_topic if it's useful. The topic you're currently working in is left out.

What the agent sees

A short block names how many records match and tells the agent the content isn't in front of it. The count includes only records that a relevance model judges a real fit for the prompt, so it is usually small, and a prompt with nothing relevant gets no count at all. The model judges only the top candidates (40 on an ordinary prompt); when every one of them fits, the number reads "N or more". If the relevance model is unavailable, the count falls back to every keyword or meaning match.

The matches are grouped by what your agent has already been given in this session:

• Never delivered: records it has never read. Up to 25 are listed with an id and a short label; the rest are counted. These are the ones to read first.
• Summary only: long records it has seen only a shortened preview of. Anything exact in them (a version, a path, a command) has to be read in full first.
• Held: records it read earlier in the session. It keeps using them, and reads them again if they've dropped out of its context.
• Changed since it was read: records that were edited or marked disputed after the agent read them. It re-reads them.

When a record your agent already read leaves circulation, the next prompt says so, whatever you asked:

• Withdrawn: the record was deleted. The agent stops relying on it.
• Contested: another record contradicts it. The agent asks you which is true and saves your answer. See settling contradictions.
• No longer current: the record was replaced. The agent reads the replacement, or works the answer out again if there is none.

As the agent reads what it needs, the block shrinks to a line or two.

How the search works

• Three kinds of record: your facts, your past work (actions), and your knowledge docs.
• Keyword and meaning together: a record matches whether you used its exact words or a paraphrase. Common words (the, with, pour) and words shorter than three letters are not searched as keywords, and a record that only shares words with your prompt must also be close to it in meaning.
• Ranked by relevance: recent, repeatedly confirmed and human-stated facts rank higher, and near-duplicates are merged.
• Linked context: when your agent reads the matching records with recall, it also gets the records directly linked to them in the knowledge graph, such as the doc behind a fact or the fix behind an incident. The per-prompt pointer counts and names only the records that matched your prompt.
• Follow-ups: your prompt is searched together with the agent's previous message, so "go ahead and do it" is searched with the subject it refers to.
• Questions about past work: prompts like "did we test this against prod?" or "what's left to ship?" widen the search, and the agent is told to check the record before saying something wasn't done. When nothing matches, it is told so.

A fact that came from a knowledge doc carries a source: line naming that doc or section. The agent reads the source before acting on anything exact in the fact. A fact with no source line is the whole record.

Size limit

Everything added to one prompt (this block, the rules names, the voice note and any colleague note) shares one size limit. Each keeps a minimum short form, so none of them disappears; when space is tight, the list of never-delivered records gets shorter first. Sizes are counted in characters, so a profile written in French or Japanese gets as much room as one in English. If even the short forms don't fit, your agent is told what was left out.

Reading the records

When the agent calls recall (or fact_search for facts only), it gets the records themselves: text, scores and sources. Only the automatic per-prompt block is a count. You can run the same search yourself on the Search page of the dashboard.

Related

• Persistent memory — what recall reads from.
• Session-start briefing — the context given before the first prompt.
• Personas & the social map — your traits and your colleagues'.
• Voice — how you want the agent to write.
• Rules — how rules reach the agent.
• The knowledge graph — the links recall follows.