Smart memory search by default

Memory recall now defaults to Smart: search the compact index, inspect context around relevant results, then fetch only the selected details. The agent reasons about relevance between these three steps. Local, cloud/Pro, and standalone-server instructions describe the tools available in that runtime.

The optional one-call lookup is now dumb_search. Its description is explicit: “Doesn’t reason about the results. Fast, but shallow. Use Smart for reliable recall.” It performs deterministic title matching and does not reason about relevance.

Migration

• Use the mem-search skill for default Smart recall. Local tools are search → timeline → get_observations; Cloud/Pro starts with memory_search and can fetch selected summaries with get_summaries.
• Replace calls to the old mem_search guided/auto API and flash_search with the Smart workflow, or explicitly choose dumb_search for a shallow lookup. Auto/guided modes and continuation cursors are removed.
• Refresh existing MCP connections so clients load the new tool catalog and descriptions.

Correctness

• Cloud timeline supports context anchored to a selected search result.
• Standalone-server recall stays scoped to its configured server project rather than reading a local worker's memory.
• Selected-detail output reports its bounds honestly and preserves selected record identities when long content must be shortened.
• Native memory lookup uses observation-only indexes, and mixed local indexes retain session/prompt ID prefixes so follow-up reads cannot select a different record kind with the same number.
• Installer prompts, public documentation, integration instructions, and generated memory guidance use the same Smart-first policy.

Other changes since v14.0.1

• Added Hermes Agent memory support and installation, including safe handling of existing YAML configuration.
• Added fallback between an Anthropic API key and the Claude subscription, with quota classification and recovery back to the primary provider at a message boundary.
• Kept worker HTTP requests responsive during local context reads, bounded broad full-text fallback searches, and removed quadratic tag-stripping stalls.
• Improved memory correctness: type="all" searches every category, deleted observations remove their Chroma documents, unexpected Chroma exits reject pending requests, and failed Windows checkpoint renames preserve the previous checkpoint.
• Tightened privacy and request handling: fully private prompts no longer become stored user requests; key redaction handles trailing dashes; internal Grok research briefs do not become user work; multilingual work requests are recognized; and OpenRouter field condensation avoids unnecessary reasoning.
• Improved installation and runtime reliability across Windows, macOS, Conda, Qoder, Goose, OpenCode, Antigravity, Codex, and TeamBrain. Fixes include launcher self-recursion and working-directory handling, stale temporary-directory cleanup, actionable authentication and missing-runtime errors, and correct host hook registration.
• Added optional cross-project context with CLAUDE_MEM_CONTEXT_PROJECTS=all, corrected folder context ordering and glob matching, expired orphaned summary spool entries, and improved effective-data-directory log lookup and cloud authentication telemetry.

Validation and delivery identities are recorded separately in the release receipt after the final builds, deployment, and package checks complete.