The Knowledge Library is a curated store of external content you want the agent to reference. Unlike Memory (which builds automatically from conversations), Knowledge is intentional: you choose what goes in.
Add files, folders, or web pages to the library. Crew chunks, indexes, and embeds them so the agent can search and reference the content during conversations. Since 0.12 you can also register an existing Amazon Bedrock Knowledge Base, which is queried live instead of ingested. Use the library for:
- Project documentation the agent should know about
- API references and specs
- Runbooks and SOPs
- Code repositories you want the agent to understand
- External pages relevant to your work
- A knowledge base your team already maintains in AWS
Open Customize -> Knowledge in the dashboard (the Knowledge & instructions group). The page has two tabs:
- Sources lists every source with its status (Active, Paused, Up to date, Error, or Awaiting confirmation), lets you add a source, pause and resume ingestion, and remove a source.
- Settings holds the library's ingestion and spend settings.
Add Source offers three kinds: a local file or folder, a Web Page, and a Bedrock Knowledge Base. Files with no extension (a README, for example) are ingested as plain text, but the file picker cannot list them, so drag and drop those.
Indexing uses Kiro requests in proportion to each source's size, and a watched folder keeps spending as files arrive. The Sources tab shows the estimated requests still needed to finish a source.
Folder sources are always deliberate. Add a folder, review the proposed source (it reports how many supported files it found), and press Start Scanning to confirm it before Crew scans anything. Crew does not create Knowledge Library folders on its own.
A folder registered automatically by an older Crew installation is held as Awaiting confirmation. It is not scanned again until you review and confirm it.
When you confirm a folder, Crew scans it recursively and keeps watching it: supported files added later are ingested on the next scan cycle (about 5 minutes). Default ignores drop OS junk files (.DS_Store, *.tmp, *.swp, and the like), a per-source ignore_patterns list narrows the scan further, and progress is shown inline as files are processed. Nothing is skipped by the pacing budget: the newest files land first and the rest continue on later sweeps.
Add Source, then Web Page, then paste an https:// link. Crew fetches the page once itself, with no agent involved, stores the text with the link and the fetch time on the source row, and fetches it again when you press Sync.
Nothing is fetched until you allow the host. knowledge.doc_ingest_hosts is empty by default; add the exact hostname first, for example:
kirocrew config set knowledge.doc_ingest_hosts '["example.com"]'
A host matches exactly (example.com does not cover docs.example.com), only https:// on the default port is fetched, the host must resolve to public addresses, and a redirect to another host is refused. A source naming an unlisted host is refused when added and when synced.
Two more sources are opt-in and off by default. knowledge.auto_ingest_artifacts indexes your saved artifacts as an "Artifacts" source kept in sync (the kinds come from knowledge.auto_ingest_artifact_kinds, default markdown, text, html and json; turning it on backfills the artifacts you already have). knowledge.auto_add_documents lets the agent add documents it reads while working to one aggregate source, shown as added by agent; the agent fetches that content with its own tools under your approval, so doc_ingest_hosts does not apply to it.
A Bedrock Knowledge Base source holds no local items. The knowledge base already keeps its own index and embeddings in your AWS account, so Crew queries it live at search time and merges the hits with local results. Each result cites the original document's link when the knowledge base records one (a crawler-built knowledge base writes it as source_uri metadata), rather than an S3 object.
What you need:
- The optional
bedrockextra: boto3 installed in the environment Kiro Crew runs in,pip install 'boto3>=1.34,<2'. Without it the add form reports the missing extra. - AWS credentials that resolve through the standard chain (a named profile, or the default credentials). Crew stores no secrets; a source carries only the knowledge base ids, the region and the profile name.
- You must be the dashboard owner. Adding and removing a Bedrock source are owner-only actions, because each search is billed to your AWS account.
To add one, open Sources -> Add Source -> Bedrock Knowledge Base and fill in:
| Field | Notes |
|---|---|
| Source name (optional) | Defaults to Bedrock KB <first id> |
| Knowledge Base IDs | Comma-separated, each a 10-character id such as KXABC12345 or its ARN arn:aws:bedrock:...:knowledge-base/KXABC12345. Up to 10 per source; repeats are dropped |
| AWS region | The region the knowledge base lives in, such as us-east-1 |
| AWS profile (optional) | A named profile from your AWS config. Leave blank to use the default credentials |
Before the first request, a card titled Confirm AWS usage before the first request shows the service ("Amazon Bedrock (knowledge base retrieval)"), the AWS account the profile resolves to, the region, and the cost note Billed per search: one Bedrock Retrieve call per Knowledge Base. Until you press Confirm and enable, no request is sent. The confirmation is recorded for that profile and region and pinned to the account they resolved to: if the profile later resolves to a different account, the confirmation is withdrawn and the card asks you to review and confirm again. Withdraw confirmation stops every Bedrock search at once.
Add Knowledge Base then runs one access check per knowledge base (one billed Retrieve each) so a missing knowledge base, a wrong region or missing permissions are refused at add time instead of failing silently later. Knowledge bases of the managed type, which reject a vector search configuration, are handled automatically.
Limits worth knowing:
- All Bedrock sources on one gateway share one profile and region. Adding a source with a different pair is refused with "Only one is supported for Bedrock Knowledge Base sources right now; remove that source first".
- Up to 64 Bedrock sources, 10 knowledge bases each.
- A search gives the remote leg about 5 seconds. If Bedrock is slow or unreachable the search returns local results only, and a result below Bedrock's relevance floor is dropped so an off-topic knowledge base does not crowd out local hits.
- A search scoped to a namespace stays local; a search scoped to one source queries only that source.
Removing a Bedrock source removes only the registration: nothing is stored locally, and the knowledge base itself is not touched. The source shows a Live badge in the Sources list.
The agent searches through local_knowledge_search when your request explicitly calls for stored documentation or matches a configured knowledge source. Search runs three legs and fuses whatever they return, so it degrades in a defined order instead of failing:
- Full: the embedding model is loaded, so meaning-based, keyword and graph matching all contribute to the ranking.
- Keyword and graph: no embedding model, so results come from full-text and graph matching. Exact wording still works; paraphrases are weaker.
- Keyword only: the graph leg finds no matching entity, so full-text matching answers alone. It needs no model and is always available.
When the embedding model is unavailable the result header, or the "No relevant knowledge found." answer, says that only keyword results were used. Registered Bedrock Knowledge Bases add a fourth, remote leg merged by rank.
Scope a search to one source or namespace when several sources cover similar terms. Use the source listing first to discover the stable identifier rather than guessing it. Each result cites where it came from: Source for a document, File for a file inside a synced folder, Artifact for something you saved, and Link for a web page.
Folder scans recognize common source files: Python, Java, TypeScript, JavaScript, Rust, Go, Ruby, PHP, C, C++, C#, Kotlin, Swift, Scala, shell, and PowerShell. Markdown, plain text, AsciiDoc, Org Mode, HTML, PDF, .docx, CSV, log files, JSON, JSONL and NDJSON, YAML, and files with no extension use their corresponding readers. .pptx is not ingested.
Files the source ignores, files beyond the configured budgets, and failed reads are reported rather than silently counted as successful; the Sources tab offers Show which files failed. The per-file cap is knowledge.max_ingest_file_mb (default 100 MB; 0 removes it).
Near-identical passages from different sources are collapsed to one result, so adding the same document directly and through a folder sync does not double every hit. A full duplicate-collapsing pass runs every knowledge.dedup_every_n_sweeps watcher sweeps (default 12), and you can run one yourself:
kirocrew knowledge dedup # preview: changes nothing kirocrew knowledge dedup --apply # collapse the duplicates
Deleting a source is permanent and cascades to everything derived from it: indexed chunks, extracted graph entities and their search contributions. The file itself stays on disk. A file that is in the library through a synced folder is removed from the Knowledge page, not from the file panel.
Read a snapshot of library counts without running a migration or repair:
kirocrew knowledge stats kirocrew knowledge stats --json
The output reports total sources, documents, and items, plus per-source counts. The agent-facing source-list tool uses the same aggregate view. Since 0.12, knowledge stats and knowledge dedup exit non-zero on failure, so scripts can trust the exit code.
| Key | Default | What it does |
|---|---|---|
knowledge.doc_ingest_hosts | [] | Exact hostnames a Web Page source may fetch. Empty fetches nothing |
knowledge.auto_ingest_artifacts | false | Index saved artifacts as a synced "Artifacts" source |
knowledge.auto_ingest_artifact_kinds | ["markdown", "text", "html", "json"] | Artifact kinds eligible for auto-ingest |
knowledge.auto_add_documents | false | Let the agent add documents it reads to one aggregate source |
knowledge.max_ingest_file_mb | 100.0 | Per-file size cap; 0 disables it |
knowledge.folder_ingest_chunk_budget | 300 | Chunks a hand-added folder may ingest per watcher sweep; 0 removes the bound, and a per-source chunk_budget overrides it |
knowledge.sweep_chunk_budget | 500 | Chunks ingested across all sources in one sweep; 0 removes the bound |
knowledge.embed_rate_limit | 120 | Embedding generations per minute across all sources; 0 removes the bound |
knowledge.dedup_every_n_sweeps | 12 | Run the full duplicate pass every Nth sweep; 0 disables |
knowledge.extraction_pool_size | 3 | Concurrent LLM workers for document extraction (1..10); applies live |
knowledge.extraction_effort | "" | Reasoning effort for the extraction pool (low .. max); empty means high. Restart required |
power.on_battery | "normal" | reduced cuts each background sweep to a quarter of the budget while on battery; paused waits for mains. Read on Linux and Windows |
| Knowledge | Memory | |
|---|---|---|
| How content gets in | You add it (files, folders, web pages, a Bedrock Knowledge Base) | Automatic from conversations |
| What it stores | External documents and references | Preferences, history, lessons, facts, episodes |
| User action | Required (add sources, confirm) | None (self-learning) |
| Best for | "The agent should know about this documentation" | "The agent should remember what I prefer" |
| Where | Customize -> Knowledge | Settings -> Memory |
Both contribute to what the agent knows, but they serve different purposes and are managed differently.
Knowledge items are embedded for semantic search using the same in-process embedding runtime as Memory. The model downloads automatically on first start (about 610 MB) without blocking the gateway. Until it lands, search falls back to the keyword and graph legs. The model can be pointed elsewhere with KIROCREW_EMBED_MODEL_URL or KIROCREW_EMBED_MODEL_PATH.
The agent searches Knowledge when:
- You ask about something that matches content in the library
- A triggered skill or workflow references library content
- The
local_knowledge_searchtool fires based on relevance signals
It doesn't dump the entire library into context: it retrieves the most relevant chunks for the current question.
Knowledge