UseMemoryMechanism — conventions for persisting important instructions

Using a mechanism to persist certain textual data between sessions is key for agentic coding. This skill assumes there is an environment variable AGENT_TECHNIQUES_MEMORY_MECHANSIM which will determine which of the guides in guides/*.md file is relevant for understanding how to use the mechanism. Use the full guide to understand the exact way to invoke or use the memory mechansim when starting, continuing, handing off, or exploring what specific work chunks need to be understood. When in doubt you can read the full narrative guide, but the high level feature sets listed below can give a way to search for the target you need within the guide.

Find or start work

Look for the starting work or claiming work items within the guide to see how to surface what are the currently slated things to work on. For example, the taskwarrior memory mechanism uses the invocation below to list the already created work chunks related to the project.

task project:<repo> list

No task yet? Create one before writing code using the memory mechanisms create functionality. For task warrior that looks like.

task add "<title>" project:<repo> +feature
task <id> start

Read prior handoffs

When starting work, look to see if there are hand off messages persisted on the work chunk in the memory mechansim or parent chunks. This involves reading the annotations or metadata on the task or work chunk. These are the last agent's (or your own past session's) notes. For taskwarrior this is

task <id> information

Log as you go

As work progresses for a particular task or chunk, create annotations about what happens like the following for taskwarrior.

task <id> annotate "decision: <choice> because <reason>"
task <id> annotate "gotcha: <non-obvious constraint>"

Handoff before finishing

Always leave one of these before context closes, even mid-task:

task <id> annotate "handoff: done=<...>, next=<...>, traps=<...>"
task <id> stop

Task fully done:

task <id> done

Reference design/plan docs in handoffs

If this task's work went through the superpowers brainstorming/writing-plans skills, they'll have written a design and/or plan doc under docs/superpowers/specs/ or docs/superpowers/plans/. Point to those paths in the handoff so a resuming agent doesn't have to rediscover them:

task <id> annotate "handoff: done=<...>, next=<...>, traps=<...>, docs=docs/superpowers/plans/<file>.md"

Convention

Prefix annotations with handoff:, decision:, or gotcha: — nothing else. This lets future agents grep the task's annotations for the kind of context they need instead of reading everything.