| tags |
agentzero |
knowledge |
indexing |
|
← Home | ↑ README
Knowledge files are documents loaded into the FAISS vector store for automatic recall. Unlike memories (which are saved by the agent), knowledge files are placed by the user into knowledge directories.
| Location |
Scope |
Notes |
usr/knowledge/ |
Global |
Always available |
usr/knowledge/main/ |
General knowledge |
Primary location |
usr/knowledge/solutions/ |
Known solutions |
Problem-solution pairs |
usr/knowledge/fragments/ |
Staging area |
Intermediate content |
<project>/.a0proj/knowledge/ |
Project-scoped |
Only when project active |
| Format |
Extension |
Notes |
| Markdown |
.md |
Primary format |
| Plain text |
.txt |
Simple documents |
| PDF |
.pdf |
Extracted and chunked |
| CSV |
.csv |
Row-based indexing |
| HTML |
.html |
Parsed text extraction |
| JSON |
.json |
String value extraction |
- Knowledge is indexed when the memory DB is initialized (first monologue of a chat)
- Files are tracked by SHA-256 checksum — only changed files are re-indexed
- The FAISS index is persisted in
usr/memory/ per subdirectory
- Re-indexing is incremental — unchanged files are skipped
# Simple: drop files into the knowledge directory
cp my-reference-doc.md /a0/usr/knowledge/main/
# The next chat session will index new/changed files automatically
| Aspect |
Knowledge |
Memory |
| Source |
User-placed files |
Agent-saved entries |
| Format |
Documents (md, pdf, etc.) |
Text snippets |
| Edit method |
File system |
memory_save / memory_delete |
| Scope |
Global or per-project |
Global or per-project |
| Recall |
Same FAISS recall mechanism |
Same FAISS recall mechanism |
- Use structured markdown with clear headings for better chunk retrieval
- Frontmatter metadata helps with filtering (
type, topic, tags)
- Keep files focused — one topic per file produces better vector matches
- Use
fragments/ as staging — consolidate into main/ when mature
- Periodic cleanup — stale knowledge degrades recall quality