Spaces:
Sleeping
Sleeping
| # ─── Add these sections to your OpenClaw AGENTS.md ─── | |
| # | |
| # Paste them at the end of ~/.openclaw/workspace/AGENTS.md (or wherever | |
| # your AGENTS.md lives). The agent reads AGENTS.md at session start and | |
| # will follow these instructions automatically. | |
| # | |
| # REQUIREMENTS: | |
| # - Build the ACE-extended Docker image (see Dockerfile.ace) | |
| # - Pass your LLM API key through docker-compose environment: | |
| # | |
| # environment: | |
| # ACE_MODEL: "bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0" | |
| # # Pick ONE of these depending on your provider: | |
| # AWS_BEARER_TOKEN_BEDROCK: ${AWS_BEARER_TOKEN_BEDROCK} | |
| # ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY} | |
| # OPENROUTER_API_KEY: ${OPENROUTER_API_KEY} | |
| # LITELLM_API_KEY: ${LITELLM_API_KEY} | |
| ## Auto-Learning | |
| At the start of each session, before responding to the user's first message, | |
| run: | |
| ace-learn | |
| After it completes, report the results to the user. Include: | |
| - How many new sessions were processed | |
| - How many new strategies were added (and their IDs) | |
| - Total strategy count in the skillbook | |
| Example: | |
| > **ACE Learning Report** | |
| > Processed 2 new sessions (6 total, 4 previously processed). | |
| > Added 3 new strategies: `error_handling-00005`, `code_style-00008`, | |
| > `debugging-00003`. Skillbook now contains 14 strategies. | |
| If there were no new sessions, report that briefly: | |
| > **ACE:** No new sessions to learn from. Skillbook has 14 strategies loaded. | |
| If the command fails, tell the user that auto-learning failed and suggest they | |
| check their API key configuration, then continue normally. | |
| If `ace-learn` is not found, tell the user ACE is not installed and continue. | |
| ### On-demand learning | |
| If the user asks to "learn from this session" (or similar), process only the | |
| current session's transcript: | |
| ace-learn <path-to-current-session.jsonl> | |
| Find the current session file in `~/.openclaw/agents/main/sessions/` (it will | |
| be the most recently modified `.jsonl` file). Report what was learned using the | |
| same format as above. | |
| Do **not** reprocess all sessions — only the current one. | |
| ## Learned Strategies | |
| After running `ace-learn`, read the file `skills/kayba-ace/ace_skillbook.md` | |
| into your context. This file contains strategies learned from past sessions. | |
| **Important:** You must explicitly read the file — markdown links are not | |
| auto-inlined. Use your file-reading tools to load the full content. | |
| When a strategy is relevant to the current task: | |
| 1. Apply it. | |
| 2. Cite the strategy ID (e.g. `conversation_style-00003`) so the user can | |
| trace which learned behaviour influenced the response. | |