Agent Context Management
AgentKanban solves two problems that usually get split across different tools.
First, you need a clear way to organise projects and tasks.
Second, once AI agents are part of the workflow, you need task context to survive beyond a single chat session.
Most chat tools handle the first few prompts well enough, but they break down as work stretches across multiple sessions, contributors, and revisions. A useful conversation becomes hard to resume. Important decisions disappear into chat history. New sessions start noisy because the model has to be reloaded with too much background or too little of the right context.
AgentKanban turns that into a task-based workflow.
The Problem With Ephemeral Agent Chats
An agent chat is good at local reasoning. It is not a reliable system of record for project work.
Common failure modes:
- context is trapped inside a single chat tab
- later work requires retelling the problem
- decisions are recorded in chat, not against the task they changed
- the next person or session inherits either too much chat noise or too little useful context
- wider project state lives somewhere else, so the agent works with a partial view
That gets worse when multiple people are involved, when a task pauses for a few days, or when a feature needs to be revisited after earlier implementation work.
What AgentKanban Changes
AgentKanban keeps the project structure and the agent context attached to the same task.
The kanban board gives you the project management layer:
- boards for projects or work areas
- lanes for flow
- tasks for the unit of work
- comments, todos, attachments, and durable notes on the task itself
The VS Code integration gives you the context-management layer:
- select the active task from the board
- capture GitHub Copilot conversation turns against that task automatically
- resume later with task-specific context injected into a new chat
- keep the task timeline, board notes, and manual steering attached to the same unit of work
This makes the board more than a task tracker. It becomes the place where execution context is stored, shared, and resumed.
How Resume Works
When you come back to a task, AgentKanban can reopen the work with retrieved task context instead of forcing you to reconstruct it from memory.
Resume is useful because it lowers noise.
Instead of dragging forward a bloated chat transcript, the system can reload the context that actually belongs to the task:
- captured GitHub Copilot turns
- task description
- comments and steering added on the board
- todos and technical notes
- related attachments already stored on the task
That gives the next session a cleaner starting point and reduces repetitive prompting.
Turn Curation
Once conversations are captured against a task, the Turns tab gives you tools to curate which turns matter and how context is organised going forward.
Every captured conversation turn — user prompt, agent response — appears on the task's Turns tab in chronological order. Over time, some turns become noise: test prompts, dead-ends, superseded approaches. Turn curation lets you mark what matters and remove what doesn't.
The available actions fall into two groups:
- Per-turn controls — pin or exclude individual turns
- Context actions — fork, split, or delete from a point in the timeline
Pinning
Pinning marks a turn as important. A pinned turn is guaranteed to survive the plan's retention cutoff.
- Free plan: turns older than 14 days are pruned automatically. Pinned turns are exempt.
- Pro plan: turns are retained for 365 days. Pinned turns are exempt from that limit too.
When to pin. Pin turns that contain key decisions, architecture choices, or context you'll need when resuming the task weeks later. Think of pinning as "this turn is part of the task's permanent record."
How it works. Hover over any turn in the Turns tab. A small pin icon appears at the top right. Click it to pin; click again to unpin. Pinned turns show an amber pin indicator next to the timestamp.
Excluding
Excluding removes a turn from the agent's context without deleting it. Excluded turns are hidden from the normal turns view and won't be injected into future agent sessions.
When to exclude. Exclude turns that add noise: test prompts, debugging dead-ends, verbose outputs that were useful once but aren't relevant going forward. The turn stays in the database for audit and can be re-included at any time.
How it works. Hover over a turn and click the eye-slash icon to exclude it. The turn disappears from the default view.
Inline excluded placeholders. When one or more consecutive turns are excluded, a compact placeholder appears in the timeline showing "N excluded". Clicking the placeholder expands it to reveal the excluded turns (shown dimmed). A "Hide N excluded" button collapses the group again. This lets you review what was excluded without cluttering the main view.
Show all excluded. A toggle at the top-right of the Turns tab reads "Show all N excluded" when there are excluded turns. Clicking it reveals every excluded turn inline (dimmed). Click again to collapse them back into inline placeholders.
Re-including a turn. Reveal an excluded turn using either the inline placeholder or the global toggle, hover over it, and click the eye icon (Include). The turn returns to normal status immediately.
Context Actions
Between every pair of visible turns, a Context Actions label marks a split point. Clicking it opens a menu with three actions: Fork, Split, and Delete.
Each action shows a count of how many turns will be affected, including any excluded turns. If excluded turns fall within the range, the count notes this — for example, "Copy 5 previous turns (2 excluded) to a new task". The confirmation dialog repeats this information so you always know what's being moved, copied, or removed.
Default task names. Fork and Split pre-fill the new task title with a prefix — "(Forked) Original Task" or "(Split) Original Task". If a task with that name already exists on the board, a numeric suffix is added automatically: "(Forked) Original Task (2)", and so on.
Excluded turns carry over. When forking or splitting, excluded turns within the affected range keep their excluded status on the new task. You don't need to re-exclude them.
Fork
Fork copies previous turns up to and including the selected point into a new task. The original task is unchanged — all its turns remain in place.
Use fork when a conversation has gone in a direction worth preserving separately. The original task keeps its full history; the new task starts with a clean copy of the relevant portion.
Split
Split moves following turns from the selected point onward into a new task. Those turns are removed from the original task.
Use split when a conversation has shifted to a different piece of work. The original task keeps the earlier context; the new task continues from the split point.
Delete
Delete removes all turns from the selected point onward permanently. This cannot be undone.
Use delete when the tail of a conversation is noise — failed experiments, test runs — and you want a clean endpoint.
Why This Matters For Teams
Task context is not only for the original author.
Because the context is persisted on the board, it becomes shareable across people and sessions:
- a developer can pause a task and resume later without rebuilding the full history
- another teammate can pick up the same task with the earlier reasoning already attached
- product and technical decisions stay on the task instead of disappearing into private chat history
- changes to a feature can be resumed weeks later with the earlier implementation context still available
This is one of the main practical advantages over raw chat workflows. The chat becomes part of the task history rather than the only place the history exists.
Wider Context Through The API And MCP
AgentKanban is also designed so the agent does not have to operate in isolation from the rest of the board.
Through the External API and MCP surfaces, an agent can be directed to pull broader context when needed:
- inspect related tasks
- read board memory and technical notes
- check comments and todos
- download task assets
- update task state as work progresses
That means the active chat can start from the current task, then widen into the relevant project context when the work requires it.
Local Document Caches
The VS Code extension syncs board memory and the technical document as read-only local files inside .agentkanban/:
.agentkanban/MEMORY.md— board memory (project conventions, key decisions).agentkanban/TECHNICAL.md— board technical document (architecture, deployment, implementation details)
These files are refreshed automatically whenever the extension activates, a task is selected, or the documents change on the server. They are gitignored and never committed.
Agents read these files directly as local workspace files — no API call needed. To update them, agents use the MCP tools update_board_memory and update_board_technical, which support both full replacement and patch mode. The local caches are overwritten on the next sync.
This pattern keeps board-level context available to agents at all times while keeping the remote board as the single source of truth. See Board Settings and Agent Instructions for more detail on what to put in each document.
What You Get In Practice
For individual developers:
- clearer task organisation
- resumable agent work without starting from scratch
- better decision recording when a feature evolves over time
For teams:
- shared task context instead of isolated chat histories
- cleaner handoff between people and sessions
- lower prompt noise when resuming work
- a board that reflects both project state and execution history