Tools for long-running agent work

Keep the context.
Match the model to the work.

MipMap keeps an ongoing local conversation useful as it grows. CLM—the Contrastive Language Model service—helps an agent choose among the models available to you and check for common mistakes in how the work gets done.

Use them together, or add CLM to Claude Code or Codex on its own.

01 · MipMap

Return to the work with the thread intact.

Imagine returning to a project after a week. The last decisions are still detailed, older work is visible as progressively smaller summaries, and the agent can open the original source messages when a detail matters. Like graphics mipmaps, the conversation is represented at different resolutions.

01

Recent context stays close

Each turn starts with a compact view of the conversation: recent messages in detail and older history at broader resolution.

02

Original messages stay reachable

The agent can zoom into summary ranges, search literal text, and inspect timestamps before relying on a remembered detail.

03

Your working history stays local

The durable log and summary tree live in your MipMap directory. Selected context still goes to the model providers that answer and summarize.

02 · CLM

Choose deliberately. Check the work.

CLM is a hosted decision service. Its integrations can route suitable work among configured options and check an agent trace for specific behavior problems. The result is a decision, not a guarantee.

A

Quick lookup

Use a lighter available option when the task is bounded.

A straightforward lookup should not need the same compute as a difficult investigation. CLM helps the agent choose a suitable available model for the task.

B

Difficult debugging

Keep stronger models available for work that needs them.

Complexity, policy and provider capacity can all affect the choice. Availability still depends on your own provider account.

C

Evidence matters

Catch a testing claim without supporting evidence.

The Claude Code and Codex integrations can flag common behavior mistakes at the end of a turn and ask the agent to verify or fix them.

Together

One thread, one task, the right amount of context.

  1. 1

    Continue the conversation

    Open the same local MipMap memory, even when the project has been quiet.

  2. 2

    Ask for a change

    Your current request becomes the task for this turn.

  3. 3

    Choose from available models

    CLM reads the current request and chooses among your available models. Your full conversation history stays out of the routing request.

  4. 4

    Recover history and work

    The selected agent receives MipMap's compact view, opens source messages when needed, and works in your repository.

Know the boundaries

Models

Your Codex or Claude account supplies agent turns. CLM chooses only among options offered to it.

Memory

MipMap stores its durable memory locally. Relevant content goes to your selected agent and summarizer.

Billing

MipMap summarization makes separately billed Anthropic API calls; it does not use a Claude subscription.