v1.9.3 — active-recall trifecta everywhere + plugin runtime rule + Tier-1 one-liner

Closes the active-recall gap across ALL delivery paths. Kit: CLAUDE.md.tmpl mandates actively calling memory_recall; install_tier2 ships the rule with the hook. PLUGIN: new recall-rule-reminder hook injects the active-recall rule each turn (a plugin can't write CLAUDE.md) + cache.py bundled = full parity. Tier 1 (bare MCP): portable AGENTS.md-first drop-in-terminal one-liner detects CLAUDE.md/GEMINI.md/AGENTS.md across ~28 tools. Honest copy: deterministic injection attempt (fail-open) + active-recall rule; rule and hook are complementary, not pick-one.

v1.9.2 — recall_verified auto-retry (hook + provider)

Client 1.9.2: the recall-first hook (both copies) and the Hermes provider now auto-retry with verified=true on abstain-with-results, injecting the cross-encoder briefing marked (verified recall) — fail-open. Routes traffic to recall_verified. Version lockstep across all manifests + KIT_VERSION 1.9.0 to 1.9.2; README seven tools + recall_verified row. Pairs with server-side routing (recall_gated + search now name recall_verified).

v1.9.0 — UltraMemory Agent Kit + one-line installer

Tier 3 — UltraMemory Agent Kit + one-line guided installer

New: the UltraMemory Agent Kit — everything in Tier 2 (recall-first hook + Turbo Token Saver client cache) plus the full operating harness for Claude Code:

• Claude Code plugin — /plugin marketplace add LogicLabsAI/ultramemory-mcp then /plugin install ultramemory-kit@ultramemory: harness skills (atomic checklists, checklist-bound execution, swarm, research), checklist-worker + checklist-verifier subagents, recall-first/capture hooks with the ejected cache, CLAUDE.md + gate templates, and an adapted playwright-human-mode skill (MIT).
• One-line guided installer — bash <(curl -fsSL https://ultramemory.io/kit.sh): Tier 2 or Tier 3, guided prompts, Playwright opt-in, idempotent, --dry-run / --non-interactive / --tier N / --uninstall, manifest-driven clean uninstall.
• CLI — ultramemory kit install|uninstall|export|check (uvx-friendly).
• MCP bundling — Context7 (keyless) preconfigured; Exa as bring-your-own-key.
• Export pipeline — scripts/export-kit.sh with a fail-closed secret deny-scan, sha256 kit manifest, and --check drift guard.
• README — new three-tier install options: Tier 1 "UltraMemory (MCP)" · Tier 2 "UltraMemory + Turbo Token Saver" · Tier 3 "UltraMemory Agent Kit".

Version lockstep 1.9.0 across pyproject.toml, plugin.yaml, server.json, and the plugin manifest. PyPI: ultramemory-hermes 1.9.0.

v1.7.0 - token-economics preview tier

Highlights (v1.4.0 → v1.7.0)

Token-economics preview tier (v1.7.0)

• Preview mode — the SDK clients (Claude Code recall hook + Hermes provider) request recall with mode: "preview": non-policy facts render as single lines under the normal section headers; [COMPANY POLICY] cards always render whole. Strictly a server opt-in — the new mode / exclude_ids params are opt-in per request, and OAuth connector defaults (claude.ai, Claude Desktop, ChatGPT) are frozen/unchanged, locked by a golden test.
• Session dedupe via exclude_ids — fact_ids already delivered this session are sent back as exclude_ids, so repeat turns spend budget on fresh facts instead of re-sending what the model already holds.
• Client cache — ~/.ultramemory/cache.json, ejected by ultramemory enable (user-editable, chmod 600, LRU-bounded at 500 entries / ~1 MB). Memoizes identical recall queries for 5 minutes (repeat query = zero HTTP calls) and tracks each session's seen fact_ids for 24 h.
• Hook budget — ULTRAMEMORY_HOOK_BUDGET (default 2000 chars ≈ ~2.8 KB/turn injection) plus ULTRAMEMORY_PREVIEW, ULTRAMEMORY_CACHE, and ULTRAMEMORY_MIN_CONFIDENCE tunables.

Gemini CLI support (v1.7.0)

• Connect over Streamable HTTP with OAuth: httpUrl block in ~/.gemini/settings.json, then /mcp auth ultramemory (bearer-header fallback for headless), plus a recall-first GEMINI.md snippet.

Recall hook + Hermes steering (v1.5.0–v1.6.0)

• Gated quiet-on-abstain recall hook with sectioned-briefing injection and latency tiers.
• Claude Code hook sends a max_characters cap sized to fit the ~10k additionalContext window; Hermes steers policy questions to recall_gated.

Capture fidelity (v1.4.0)

• Full-turn capture (including tool output), whole-session rollup card, wayback-grade session snapshot Skill.

Full changelog: v1.3.0...v1.7.0