Knowledge Bases

The memory bank grows one learned entry at a time. Knowledge bases solve the opposite problem: you already have a body of written material, and you want your agents answering from it today. Import a document and every paragraph becomes a searchable, citable chunk. The two stores are complementary by design, and agents search them together.

How Import Works

When a document enters a knowledge base it is split into paragraph-sized chunks, and each chunk is embedded by the memory bank service, the same vector treatment conversation turns and memories receive. Chunking at the paragraph level is the practical sweet spot: each piece is small enough that retrieval is precise, and large enough to carry a complete thought, so the chunk an agent retrieves usually contains the full answer rather than a fragment of one.

You can keep any number of knowledge bases, one per product, one for policies, one for technical reference, whatever shape your material has. Each base is named, listed, and managed on its own, so retiring an outdated manual never touches the rest.

Searching

Agents reach knowledge bases two ways. The dedicated knowledge search queries a specific base when the agent knows where the answer lives. The wider query all searches every knowledge base and every memory type together in one call, the broadest question the system can ask itself, and the standard opening move when an agent faces an unfamiliar question:

php tools/memory_query.php query all "search text"
php tools/memory_query.php knowledge search "base name" "search text"
php tools/memory_query.php knowledge list | info | create | add | delete

Because search is by meaning, phrasing does not have to match. A customer asking "can I get my money back" finds the refund policy paragraph even though the document never uses those words.

Which Store Should Hold It

A simple decision rule keeps the two stores clean. Material that exists as documents, manuals, policy pages, price lists, spec sheets, belongs in a knowledge base: it arrives in bulk, it has an author, and it gets replaced wholesale when a new version ships. Things the system discovers in the course of working, a customer's preference, a fix that worked, the right way to call an API, belong in the memory bank: they arrive one at a time and accumulate. When you are unsure, ask where the update will come from. If a future version of a document will replace it, knowledge base. If experience will refine it, memory.

Writing for retrieval is worth a minute of care on import. Documents whose paragraphs each make one self-contained point retrieve beautifully, because the chunk boundary is the thought boundary. Long run-on sections retrieve less precisely, so if a document matters, breaking its wall-of-text sections into focused paragraphs before import measurably improves the answers agents give from it.

The Backbone of Verified Support

Knowledge bases carry their biggest responsibility in customer service. The customer service preset answers a question only when the knowledge base contains a verified answer for it, which is the platform's standing guarantee against improvised support replies. When a question arrives that the knowledge base cannot answer, the agent raises a flag instead of guessing.

That flag starts the growth loop. You answer the question once, in the UI or by telling the master agent, and the answer joins the knowledge base. The next customer who asks gets the verified answer automatically. Over a few weeks of real traffic, this loop builds exactly the knowledge base your customers actually need, ordered by what they actually ask, with every entry written or approved by you.

Fed by Research

The research pipeline can write into knowledge bases too. Set kb_insert in the pipeline settings with a target kb_name, and each completed research task deposits its verified findings as knowledge base content. The combination is powerful: ask for research on a topic your support team keeps meeting, and the verified, source-checked result becomes answer material your agents can cite the same day.

Managing Your Bases

The Knowledge tab in the web UI is the management surface: create bases, import documents, browse chunks, search interactively, and remove what no longer applies. The CLI offers the same operations for scripted workflows, and because knowledge bases live in the same local database as everything else, they ride along in your ordinary backups and they work identically on the local backend and on Adaptive Recall.

The habit that pays off most: import what you already have on day one. Product docs, your policy pages, your best support replies, anything a competent new employee would be handed. Agents query before acting, so every paragraph you import is immediately part of how the whole system answers.