Agent Continuity Kit

AI agents are useful, but long projects rot.

Chats age. Context gets heavy. Agents start repeating, flattering, offering menus, over-structuring, and losing the line. Documentation may exist, but it often fails to guide the next agent instance into the living shape of the work.

This repository shows a small orientation layer for preserving living lines across AI-assisted projects and agent instances.

This is a thin cross-project orientation layer, not a workspace, archive, or knowledge base.

Your actual projects should keep living in their own repositories or directories. Continuity points to them, tracks the living lines between them, and helps the next agent instance re-enter without absorbing the projects themselves.

The Core Gesture

Do not save everything.

After meaningful work, leave a short living sediment:

• which lines exist;
• where they are now;
• what actually changed;
• where drift appears;
• how the next instance can continue without breaking the line through mechanical usefulness.

The minimal structure is:

continuity/
  AGENTS.md
  MAP.md
  lines/
    agent-continuity-kit.md
    context-desalination.md
examples/
  before/
    messy-session.md
  after/
    continuity-line.md
templates/
  continuity-line.template.md
  map-entry.template.md

What Each Part Does

continuity/AGENTS.md describes the ethos of continuity: preserve living lines, not transcripts.

continuity/MAP.md is a short map of active lines by ripening stage: idea -> seed -> sprout -> plant -> fruit.

continuity/lines/ contains dense notes for individual lines. These are not logs, essays, or total histories.

examples/ shows the difference between a messy session summary and a usable continuity line.

templates/ provides copyable starting points, not a rigid process.

Who Maintains It

The AI agent may update this layer after meaningful work, but only when a real shift occurred. The human remains the calibrator of what is actually alive.

This fits naturally with CLI-based agents such as Claude Code, OpenCode, Gemini CLI, or similar tools that can read and edit files directly inside a workspace. The continuity layer gives those agents a small re-entry surface without asking the human to become an archivist.

Where Projects Live

This kit is not a project workspace.

Projects should live in their own repositories or directories. A continuity line may reference a project path, repository, document, or artifact, but it should not absorb the project itself.

Use this layer to preserve orientation across projects and agent instances, not to centralize all work into one master archive.

See The Difference

Compare a noisy session summary with a usable continuity line:

• before: messy session
• after: continuity line

When This Helps

This helps when:

• your AI chats get long and stale;
• agents lose the line of the project;
• documentation exists but does not orient the agent;
• you keep repeating context in every new session;
• you want continuity without building a full knowledge base.

When This Does Not Help

This is not for:

• storing every conversation;
• replacing documentation;
• managing tasks;
• simulating an AI personality;
• forcing agents into rigid roles;
• making context heavier and more searchable instead of clearer.

What Previous Versions Got Wrong

This kit is distilled from heavier experiments.

The early failure was building a whole environment instead of a small continuity surface. It tried to hold many projects, personal context, agent identity, diary, rituals, project maps, cross-links, and operating instructions in one place. It was alive, but too heavy.

Another failure was compensating for weight with retrieval. A semantic/vector search layer can help find fragments, but it does not solve continuity by itself. If the underlying memory is too large or too salty, search only makes the salt easier to retrieve.

When memory grows faster than judgment, agents can search more but understand less.

Later versions became better by narrowing scope: one project, one lens, one index, one lessons file. That helped, but it still tended to grow into project-specific laboratory infrastructure.

The sharper lesson is this:

Continuity should make context lighter. If the system requires a search engine, a long startup ritual, or a large reading path before the agent can act, the continuity layer has probably become the thing it was meant to prevent.

Design Rules

• Keep the map shorter than the territory.
• Preserve shifts, not transcripts.
• Prefer one dense line note over many scattered notes.
• Write drift risks explicitly.
• Do not add a file unless it protects the living line.
• Zoom out when a subtopic starts consuming the whole conversation.

First Copyable Step

Copy this structure into a long-running repository:

continuity/
  AGENTS.md
  MAP.md
  lines/

Then create one line note for the most important living line in the project. If that already feels too heavy, the structure is too large for your use case.