Recall

Read the memory back, packed to a token budget, every row naming the source it came from.

• Retrieval

  For every read an agent makes. Every read here is packed to a token budget you set and every row carries the ids it came from, so an answer built on it can cite its source rather than assert it.

• Agent sessions

  For an open-ended goal. A run is a durable session bound to one project: you open it once, append messages, and read the answer back as a single JSON read or as a live stream — the agent decomposes the goal, retrieves repeatedly, and names the declarations each claim rested on. Nothing on this path mutates a codebase: when a turn concludes a change is needed it writes a plan — the reasoning, the proposed changes, the terms it read without changing — and stops. Applying it is a separate, explicit approval on the same run, which is what makes the surface safe to put behind an editor's "apply" button.