Context Budgeting

**Quick judgment**: A solid, opinionated framework for managing the context window in OpenClaw agents. It is most useful for long-lived, tool-heavy sessions where context routinely nears capacity and naive truncation causes task confusion. **What it does well** - Structures context into clear tiers (objective, short-term history, decision logs, background), making it easier to decide what to keep vs. summarize or evict. - Enforces a **pre-compression checkpoint** into `memory/hot/HOT_MEMORY.md` (status, key decisions, next step) before any cleanup, reducing the chance that compaction destroys the working set of the task. - Hooks into **Heartbeat** as a GC trigger when context usage exceeds ~80%, and encourages deleting large raw artifacts once summarized, improving both cost and latency. **Key risks / limitations** - Requires disciplined implementation: if the checkpointing step or `gc_and_checkpoint.sh` invocation is skipped or misconfigured, important state can still be lost during compaction. - Quality of summaries and decision logs is critical; poor summarization will still lead to “memory loss,” just in a more structured way. - Assumes specific file paths and scripts (e.g., `skills/context-budgeting/scripts/gc_and_checkpoint.sh`, `memory/hot/HOT_MEMORY.md`); projects that don’t follow this layout need adaptation. - Adds operational complexity (heartbeat integration, script management) that may be overkill for short or simple sessions. **Recommended scenarios** - Long-running, stateful agents that frequently approach or exceed 80% of their context budget. - Workflows that generate large, transient outputs from tools or APIs and need systematic cleanup after summarization. - Deployments that care strongly about token cost and latency and are willing to invest in structured context lifecycle management. - Teams that want a repeatable, auditable procedure for compaction instead of ad-hoc truncation of conversation history.