Memory and code search

Agents waste most of their effort rediscovering things: where a behaviour lives in the code, why a decision was made, what went wrong last week. I gave every agent in my estate the same memory, served by FAFOHive. These eight tools are how an agent asks it questions. Code is indexed by syntax tree, so a search returns whole functions and types with file and line. The code graph answers structural questions without a model. Decisions and fixes are kept as observations, so the reasoning behind a change outlives the session that made it.

The tools

list_indexes lists the code and reference collections that exist and how much each holds. Its names are the scope every search uses.

search_code finds the implementation behind a behaviour in an indexed codebase and returns ranked code chunks with file paths and lines. A search can cover the whole codebase, one directory, or a chosen set of files that work together, and it has an exact-symbol mode for named functions.

search_references searches external material: documents, library source and crawled documentation. It is where an agent looks for a reference implementation or a specification before writing its own.

trace_symbol_dependencies walks the code graph around a function, type or trait to show callers, callees, imports and implementations, with a chosen direction and depth. No model is involved, so it answers fast and it cannot make an edge up.

explain_code_path turns a path through the code graph into a readable explanation of how execution or data flows between two points. The narrative is built only from edges the graph contains, which is the subject of a code explainer that is not allowed to invent.

search_observations searches recorded decisions, discoveries and fixes by meaning, by keyword or by both, with project and date filters.

create_observation records a decision, fix or discovery with a title, a narrative of what happened and why, and optional tags and affected files.

search_agent_history reads the raw activity log of agent runs with exact filters on job, trace, event, model and time. It reconstructs one particular run, where observation search finds lessons across many.

Eight tools is a small surface on purpose. Every agent connects to it, and the less it has to learn, the more often it uses memory before it guesses.