Memory

One query surface, not one store per agent#

Every agent that needs to know something asks the same layered search. It merges the project-local index, the learning tiers, the manifest facts a scan produced, and any live research — and returns one ranked list.

Each result carries the source it came from, so the prompt underneath can tell a proven skill from a past gotcha from a known architectural fact. That distinction is the difference between memory that helps and memory that misleads.

What gets remembered#

Code

Indexed chunks of the repository itself, so a question about the codebase is answered from the codebase.

Decisions

What was chosen and why, so a later run does not quietly reverse it.

Plan summaries

What a previous run set out to do and what came of it.

Gotchas

The specific traps this repository has — the things that are only obvious after they have cost someone an afternoon.

Skills

A repair or approach that demonstrably worked, promoted after a confident success and deduplicated against what already exists.

Reflections and failure patterns

What was attempted and did not work. Knowing the dead ends is worth as much as knowing the routes.

How results are ranked#

Relevance to the query, recency, and the historical success rate of that piece of knowledge — combined, not chosen between. A fact that was true once and has failed twice since ranks below one that keeps working.

That last term is what makes the memory improve rather than merely accumulate. Storage is easy; deciding what deserves to surface is the part that matters.