Release v1.7.0 - token-economics preview tier · LogicLabsAI/ultramemory-mcp

Highlights (v1.4.0 → 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.

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.

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.

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

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