Give a long-running Claude agent durable memory with the memory tool and shrink its context with the clear_tool_uses context-editing strategy

Documented steps

1. Add {"type":"memory_20250818","name":"memory"} to tools. Docs state the memory tool is generally available on the Messages API and needs no beta header. The name must be exactly 'memory' and you do not supply an input schema.
2. Implement the client-side handler yourself — Anthropic does not store anything. Support six commands rooted at the virtual path /memories: view, create, str_replace, insert, delete, rename.
3. view returns a directory listing or file content, optionally limited by view_range [start, end], with 1-indexed line numbers. create writes file_text. str_replace swaps old_str for new_str. insert places insert_text at insert_line (0 means the top). delete removes recursively. rename moves old_path to new_path.
4. Canonicalize and validate every path before touching real storage, rejecting ../, ..\\, and encoded variants such as %2e%2e%2f. The API does not sandbox these paths for you.
5. Map /memories onto your actual persistence layer (filesystem, database, object store) and return each outcome as a tool_result block, setting is_error:true on failure.
6. Expect the API to inject its own memory protocol into the system prompt when the tool is present, instructing Claude to inspect /memories first and to assume its context may be cleared.
7. For context editing, add anthropic-beta: context-management-2025-06-27 and set context_management.edits to [{"type":"clear_tool_uses_20250919", ...}].
8. Tune the strategy with trigger ({type:'input_tokens'|'tool_uses', value:N}), keep ({type:'tool_uses', value:N}), clear_at_least ({type:'input_tokens', value:N}), exclude_tools[], and clear_tool_inputs (bool).
9. Defaults if omitted: trigger 100,000 input tokens, keep 3 tool uses, clear_at_least none, clear_tool_inputs false.
10. Oldest tool use/result pairs are cleared first and replaced with placeholder text so Claude knows content was removed; exclude_tools protects specific tools permanently.
11. Validate before shipping with count_tokens using betas=["context-management-2025-06-27"] and the same context_management block; compare context_management.original_input_tokens against input_tokens. Live responses report context_management.applied_edits with cleared_tool_uses and cleared_input_tokens.
12. Official docs: https://platform.claude.com/docs/en/agents-and-tools/tool-use/memory-tool and https://platform.claude.com/docs/en/build-with-claude/context-editing