THE LABS BUILD MEMORY
FOR THEIR OWN TOOL.
Cursor remembers Cursor. Copilot remembers Copilot. Claude remembers Claude. Your team needs one record that survives the tool and the model — every agent that speaks MCP.
FAQ
- Where do AI coding tokens actually go?
- In agentic software engineering studies, most spend is not “write the first file.” Tokenomics (arXiv:2601.14470) finds input tokens ~54% of consumption on average, and iterative code review ~59% of tokens — refinement and verification dominate. Scout’s job is not to invent a savings %. It keeps one team record and projects only what’s current into the session bundle so agents stop re-pasting stale team state as input.
- Why does “shared context” get expensive across a team?
- Ten developers, ten agents, ten versions of what happened today means the same decisions and interface changes get re-explained into prompts. Industry work (AgentDiet, arXiv:2509.23586) shows agent trajectories fill with useless, redundant, and expired information. Scout is the opposite of stuffing more transcript into the model: one record, every agent reads and writes it, and only the current page is what the model sees. (We do not call ourselves a “shared context layer.”)
- Is Scout a token compressor or summarizer?
- No. Compressors and trajectory reducers (e.g. AgentDiet) delete waste inside one agent’s run. Scout is the team’s append-only record plus a current-state projection over MCP — rules, contracts, handoffs, blockers — sized to a token budget you set at session start. Different problem. Cite AgentDiet only to show waste is real industry-wide, not as our method.
- If we send fewer tokens, do we always pay less?
- Not always. Provider caching and pricing mean token count ↓ is not a guaranteed dollar ↓. Treat industry token shares (Tokenomics) as a map of where cost concentrates — input and review loops — then design the workflow so the model gets a current, budgeted bundle instead of a growing paste of last week’s state.
- How does Scout keep the session bundle from turning into last year’s folder?
- Everything is kept (append-only). The model does not read the whole history. Scout projects what is currently true — newest standing facts, open contracts, handoffs, rules — into one session-start call within your token budget. That is the mechanism. No Scout performance %.