v2.1.2 focuses on operational trust: exact search stays exact, purge really removes content, contradictions are directly inspectable, and the update flow no longer depends on copied curl commands.
### Added
- **Concrete search mode** — `search` now auto-detects literal queries such as quoted strings, env vars, UUIDs, paths, and code identifiers. Those queries take a keyword/literal path that skips HyDE, semantic fusion, FSRS reweighting, retrieval competition, and spreading activation so exact matches land first.
- **Irreversible purge** — `memory(action="purge", confirm=true)` permanently removes memory content and embeddings, scrubs `insights.source_memories`, detaches temporal-summary children, prunes graph edges, and writes only a non-content `deletion_tombstones` row for sync/audit.
- **First-class contradictions tool** — new `contradictions` MCP tool scans a topic or recent memories for trust-weighted disagreements using the same local contradiction logic as `deep_reference`.
- **Simple update flow** — `vestige update` refreshes the installed binary and companion Sandwich files without requiring users to paste a curl installer.
- **Pro waitlist preview** — `/dashboard/waitlist` adds a local-only marketing surface for Solo Pro and Team Pro early-access signups. `VITE_WAITLIST_ENDPOINT` and `VITE_SUPPORT_BOT_ENDPOINT` are opt-in dashboard env vars, so no signup data is captured unless endpoints are configured.
- **Dream connection persistence cap** — dense single-domain dreams now persist every connection discovered in that run instead of losing everything beyond the old 1,000-entry live buffer. The live dreamer buffer now keeps up to 200,000 high-scoring recent connections, and the MCP `dream` tool exposes `min_similarity` for corpus-specific tuning.
- **Embedding-model upgrade repair** — `vestige consolidate` now re-embeds every missing or active-model-mismatched memory in one pass, so v1/v2 mixed stores are no longer left partially unreachable after only the first 100 legacy embeddings are regenerated.
- **Merge-safe imports** — `vestige portable-import --merge` can merge into non-empty databases. It applies `sync_tombstones`, keeps newer local memory rows on timestamp conflicts, preserves stable IDs, and rebuilds FTS after import.
- **File-backed two-way sync** — `vestige sync <archive>` performs pull-merge-push through a shared portable archive. This works today with Dropbox, iCloud Drive, Syncthing, Git, network shares, and shared folders.
- **Pluggable portable-sync backend trait** — core now exposes `PortableSyncBackend`, `FilePortableSyncBackend`, and `PortableSyncReport`, so non-file backends can reuse the same merge semantics without reimplementing conflict handling.
- **Portable restore merge mode** — the MCP `restore` tool accepts `merge: true` for portable archives and returns inserted/updated/deleted/skipped/conflict counts.
- **Qwen3 embedding opt-in** — build-time and runtime support for Qwen3 embeddings, with model-aware retrieval safeguards so mixed embedding models are not compared in the same vector path.
### Fixed
- **Sanhedrin, preflight, and all Vestige Claude Code hooks are optional again.** The Cognitive Sandwich installer now activates no hooks by default and leaves every preflight hook, every Stop hook, the MLX launchd service, and the 19 GB Qwen model path behind explicit `--enable-preflight`, `--enable-sanhedrin`, or `--with-launchd` flags.
- **x86-friendly Sanhedrin path.** The verifier bridge now accepts any OpenAI-compatible chat endpoint via `VESTIGE_SANHEDRIN_ENDPOINT` and `VESTIGE_SANHEDRIN_MODEL`, so Linux and Intel Mac users can opt in without MLX or Apple Silicon.
### Verified
-`cargo test -p vestige-core portable --no-fail-fast`
-`cargo test -p vestige-mcp portable --no-fail-fast`
-`cargo test --workspace --no-fail-fast`
- Installer shell/Python/JSON validation and default/preflight/Sanhedrin migration dry-runs.
v2.1.0 ships the Cognitive Sandwich hook harness for Claude Code, with a hotfix that makes every hook layer opt-in. The default installer stages files, removes old Vestige hook wiring, starts no MLX service, downloads no model, and makes no automatic model calls. Users can opt into preflight context, Sanhedrin verification, or the Apple Silicon MLX backend separately.
- **`hooks/`** — first-class harness-side companion to the Vestige MCP server. 9 production hooks designed for `~/.claude/hooks/`:
-`sanhedrin.sh` — Stop hook that invokes the local Qwen Executioner via the Python bridge.
-`sanhedrin-local.py` — local backend that POSTs to `mlx_lm.server` (`localhost:8080`) with Vestige evidence injected via the dashboard `/api/deep_reference` HTTP endpoint. TRUST_FLOOR=0.55 evidence filter + topical-relevance gate + inference-verb ban + 8 worked few-shots covering true positives AND false-positive guards.
-`synthesis-preflight.sh` — UserPromptSubmit hook that POSTs the user prompt to `/api/deep_reference` and injects the trust-scored reasoning chain into context.
-`cwd-state-injector.sh` — captures git status, branch, modified files, open PRs/issues.
-`vestige-pulse-daemon.sh` — surfaces fresh Vestige dream insights from the past 20 min.
-`preflight-swarm.sh` — spawns the `lateral-thinker` subagent in fresh context for cross-disciplinary structural parallels.
-`synthesis-stop-validator.sh` — Stop hook regex against forbidden hedging patterns.
-`veto-detector.sh` — fast 50ms regex pre-screen against `veto`-tagged Vestige memories.
-`synthesis-gate.sh` — legacy v1 trigger (kept for backward compat).
- **Dashboard `/api/changelog` endpoint** — bounded REST event feed for recent `DreamCompleted` and `ConnectionDiscovered` events, used by the Pulse hook to inject fresh synthesis into Claude Code context.
- **`launchd/com.vestige.mlx-server.plist.template`** — auto-start `mlx_lm.server` with the Qwen3.6-35B-A3B-4bit model on login. Templated with `__HOME__` and `__MODEL__` placeholders.
- **`scripts/install-sandwich.sh`** — one-command installer that stages hooks and agents, removes old Vestige hook wiring by default, and only wires optional layers with `--enable-preflight`, `--enable-sanhedrin`, or `--with-launchd`. Backs up `settings.json` to `.bak.pre-sandwich`. Supports `--force`, `--include-memory-loader`, and `--src=PATH`.
- **`scripts/check-sandwich-prereqs.sh`** — default verifier confirms no Vestige Claude Code hooks are wired. Optional `--preflight` and `--sanhedrin` modes check the corresponding enabled layer.
- **Sanhedrin Executioner backend can run locally or remotely when explicitly enabled.** The bridge targets an OpenAI-compatible chat endpoint, with local `mlx_lm.server` + Qwen3.6-35B-A3B-4bit available behind `--with-launchd` on Apple Silicon. Anthropic API key no longer required for the post-cognitive layer. The `executioner.md` agent definition is retained as manual/fallback only when invoked explicitly via `Task(subagent_type='executioner')`.
- **All hooks sanitized for public release** — replaced hardcoded personal absolute paths with `$HOME` / `$VESTIGE_*` env vars; removed personal regex tokens.
- **NPM binary installer now follows package version** — `vestige-mcp-server@2.1.0` downloads release assets from `v2.1.0` instead of a stale hardcoded binary tag, while local workspace installs skip the release-asset download before the tag exists.
None required for existing Vestige users. The Cognitive Sandwich is opt-in via `scripts/install-sandwich.sh`; running the default installer removes old Vestige hook wiring and leaves preflight, Sanhedrin, launchd, and the 19 GB model path disabled. The MCP server, schema, and tool surface are bit-identical to v2.0.9.
Autopilot flips Vestige from passive memory library to self-managing cognitive surface. A single supervised backend task subscribes to the 20-event WebSocket bus and routes live events into the cognitive engine — 14 previously dormant primitives (synaptic tagging, predictive memory, activation spread, prospective polling, auto-consolidation, Rac1 cascade emission) now fire without any MCP tool call. Shipped alongside a 3,091-LOC orphan-code cleanup of the v1.0 tool surface. **No schema changes, tool surface unchanged (24 tools), fully backward compatible with v2.0.8 databases. Opt-out via `VESTIGE_AUTOPILOT_ENABLED=0`.**
### Added
- **Autopilot event subscriber** (`crates/vestige-mcp/src/autopilot.rs`) — two supervised tokio tasks spawned at startup. The event subscriber consumes a `broadcast::Receiver<VestigeEvent>` and routes six event classes:
-`SearchPerformed` → `predictive_memory.record_query()` + top-10 access records. The speculative-retrieval model now warms without waiting for an explicit `predict` call.
-`MemoryPromoted` → `activation_network.activate(id, 0.3)` — a small reinforcement ripple through the graph.
-`MemorySuppressed` → re-emits the previously-declared-never-emitted `Rac1CascadeSwept` event so the dashboard's cascade wave actually renders.
-`ImportanceScored` with `composite_score > 0.85` AND a stored `memory_id` → auto-promote + re-emit `MemoryPromoted`.
-`Heartbeat` with `memory_count > 700` → rate-limited `find_duplicates` sweep (6h cooldown + in-flight `JoinHandle` guard against concurrent scans on large DBs).
The engine mutex is acquired only synchronously per handler and never held across `.await`, so MCP tool dispatch is never starved. A 60-second `tokio::interval` separately polls `prospective_memory.check_triggers(Context)` — matched intentions log at `info!` level today; v2.1 "Autonomic" will surface them mid-conversation.
- **Panic-resilient supervisors** — both background tasks run inside an outer supervisor loop. If a cognitive hook panics on one bad memory, the supervisor catches `JoinError::is_panic()`, logs the panic, sleeps 5 s, and respawns the inner task. Turns a permanent silent-failure mode into a transient hiccup.
- **`VESTIGE_AUTOPILOT_ENABLED=0` opt-out** — v2.0.8 users who want the passive-library contract can disable Autopilot entirely. Values `{0, false, no, off}` early-return before any task spawns; anything else (unset, `1`, `true`) enables the default v2.0.9 behavior.
- **`ImportanceScored.memory_id: Option<String>`** — new optional field on the event variant (`#[serde(default)]`, backward-compatible) so Autopilot's auto-promote path can target a stored memory. Existing emit sites pass `None`.
### Changed
- **3,091 LOC of orphan tool code removed** from `crates/vestige-mcp/src/tools/` — nine v1.0 modules (`checkpoint`, `codebase`, `consolidate`, `ingest`, `intentions`, `knowledge`, `recall`, plus two internal helpers) superseded by the `*_unified` / `maintenance::*` replacements shipped in v2.0.5. Each module verified for zero non-test callers before removal. Tool surface unchanged — all 24 tools continue to work via the unified dispatchers.
- **Ghost environment-variable documentation scrubbed** — three docs listed env vars (`VESTIGE_DATA_DIR`, `VESTIGE_LOG_LEVEL`, a `VESTIGE_API_KEY` typo) that never existed in any shipping Rust code. Replaced with the real variables the binary actually reads.
### Fixed
- **Dedup-sweep race on large databases** — the `Heartbeat`-triggered `find_duplicates` sweep previously set its cooldown timestamp BEFORE spawning the async scan, which on 100k+ memory DBs (where a sweep can exceed the 6h cooldown) allowed two concurrent scans. Rewritten to track the in-flight `JoinHandle` via `DedupSweepState::is_running()` — the next Heartbeat skips if the previous sweep is still live.
### Verified
-`cargo test --workspace --no-fail-fast`: **1,223 passing, 0 failed**.
None. v2.0.9 is a pure backend upgrade — tool surface, JSON-RPC schema, storage schema, and CLI flags are bit-identical to v2.0.8. Existing databases open without any migration step. The only behavior change is the Autopilot task running in the background, which is `VESTIGE_AUTOPILOT_ENABLED=0`-gated if you want the old passive-library contract.
The Pulse release wires the dashboard through to the cognitive engine. Eight new dashboard surfaces expose `deep_reference`, `find_duplicates`, `dream`, FSRS scheduling, 4-channel importance, spreading activation, contradiction arcs, and cross-project pattern transfer — every one of them was MCP-only before. Intel Mac is back on the supported list (Microsoft deprecated x86_64 macOS ONNX Runtime prebuilts; we link dynamically against a Homebrew `onnxruntime` instead). Reasoning Theater, Pulse InsightToast, and the Memory Birth Ritual all ship. No schema migrations.
### Added
- **Reasoning Theater (`/reasoning`)** — Cmd+K Ask palette over the 8-stage `deep_reference` cognitive pipeline: hybrid retrieval → cross-encoder rerank → spreading activation → FSRS-6 trust scoring → temporal supersession → trust-weighted contradiction analysis → relation assessment → template reasoning chain. Every query returns a pre-built reasoning block with evidence cards, confidence meter, contradiction geodesic arcs, superseded-memory lineage, and an evolution timeline. **Zero LLM calls, 100% local.** New HTTP surface `POST /api/deep_reference` wraps `crate::tools::cross_reference::execute`; new WebSocket event `DeepReferenceCompleted` carries primary / supporting / contradicting memory IDs for downstream graph animation.
- **Pulse InsightToast (v2.2 Pulse)** — real-time toast stack that surfaces `DreamCompleted`, `ConsolidationCompleted`, `ConnectionDiscovered`, `MemoryPromoted`/`Demoted`/`Suppressed`, `MemoryUnsuppressed`, `Rac1CascadeSwept` events the moment they fire. Rate-limited to 1 per 1500ms on connection-discovery cascades. Auto-dismiss after 5-6s, click-to-dismiss, progress bar. Bottom-right on desktop, top-center on mobile.
- **Memory Birth Ritual (v2.3 Terrarium)** — new memories materialize in the 3D graph on every `MemoryCreated` event: elastic scale-in from a camera-relative cosmic center, quadratic Bezier flight path, glow sprite fades in frames 5-10, label fades in at frame 40, Newton's Cradle docking recoil. 60-frame sequence, zero-alloc math, camera-relative so the birth point stays visible at every zoom level.
- **3D Graph brightness system** — auto distance-compensated node brightness (1.0× at camera <60u,upto2.4×atfarzoom)sonodesdon'tdisappearintoexponentialfogatzoom-out.User-facingbrightnesssliderinthegraphtoolbar(☀icon,range0.5×-2.5×,localStorage-persistedunder`vestige:graph:brightness`).Composeswiththeautoboost;opacity+glowhalo+edgeweighttrackthecombinedmultipliersonodesstaycoherent.
- **Intel Mac (`x86_64-apple-darwin`) support restored** via the `ort-dynamic` Cargo feature + Homebrew-installed ONNX Runtime. Microsoft is discontinuing x86_64 macOS prebuilts after ONNX Runtime v1.23.0 so `ort-sys` will never ship one for Intel; the dynamic-link path sidesteps that entirely. Install: `brew install onnxruntime` then `ORT_DYLIB_PATH=$(brew --prefix onnxruntime)/lib/libonnxruntime.dylib`. Full guide bundled in the Intel Mac tarball as `INSTALL-INTEL-MAC.md`. **Closes #41.**
- **Graph default-load fallback** — when the newest memory has zero edges (freshly saved, hasn't accumulated connections yet), `GET /api/graph` silently retries with `sort=connected` so the landing view shows real context instead of a lonely orb. Applies only to default loads; explicit `query` / `center_id` requests are honored as-is. Fires on both backend and client.
### Fixed
- **Contradiction-detection false positives** — adjacent-domain memories are no longer flagged as conflicts just because both contain the word "trust" or "fixed." Four thresholds tightened: `NEGATION_PAIRS` drops the `("not ", "")` + `("no longer", "")` wildcard sentinels; `appears_contradictory` shared-words floor 2 → 4 and correction-signal gating now requires ≥6 shared words + asymmetric presence (one memory carries the signal, the other doesn't); `assess_relation` topic-similarity floor raised 0.15 → 0.55; Stage 5 pairwise contradiction overlap floor 0.15 → 0.4. On an FSRS-6 query this collapses false contradictions from 12 → 0 without regressing the two legitimate contradiction test cases.
- **Primary-memory selection on `deep_reference`** — previously the reasoning chain picked via `max_by(trust)` and the recommended-answer card picked via `max_by(composite)`, so the chain and citation disagreed on the same query. Unified behind a shared composite (50% hybrid-search relevance + 20% FSRS-6 trust + 30% query-topic-term match fraction) with a hard topic-term filter: a memory cannot be primary unless its content contains at least one substantive query term. Three-tier fallback (on-topic + relevant → on-topic any → all non-superseded) so sparse corpora never starve. Closes the class of bug where high-trust off-topic memories won queries against the actual subject.
- **Reasoning page information hierarchy** — reasoning chain renders first as the hero (confidence-tinted border glow, inline metadata), then confidence meter + Primary Source citation card, then Cognitive Pipeline visualization, then evidence grid. "Template Reasoning" relabelled "Reasoning"; "Recommended Answer" relabelled "Primary Source" (it's a cited memory, not the conclusion — the chain is the conclusion).
### Changed
- **CI + release workflows** — `release-build` now runs on pull requests too so Intel Mac / aarch64-darwin / Linux / Windows regressions surface before merge. `x86_64-apple-darwin` back in both `ci.yml` and `release.yml` matrices with `cargo_flags: "--no-default-features --features ort-dynamic,vector-search"`. Intel Mac tarball bundles `docs/INSTALL-INTEL-MAC.md` alongside the binaries.
- **Cargo feature split** — `embeddings` is now code-only (fastembed dep + hf-hub + image-models). New `ort-download` feature carries the prebuilt backend (the historical default); `ort-dynamic` transitively enables `embeddings` so the 27 `#[cfg(feature = "embeddings")]` gates stay active when users swap backends. Default set `["embeddings", "ort-download", "vector-search", "bundled-sqlite"]` — identical behavior for every existing consumer.
- **Platform availability in README** — macOS Apple Silicon + Intel + Linux x86_64 + Windows x86_64 all shipped as prebuilts. Intel Mac needs `brew install onnxruntime` as a one-time prereq.
### Docs
- New `docs/INSTALL-INTEL-MAC.md` with the Homebrew prereq, binary install, source build, troubleshooting, and the v2.1 `ort-candle` migration plan.
- README Intel Mac section rewritten with the working install recipe + platform table updated.
### Migration
None. Additive features and bug fixes only. No schema changes, no breaking API changes, no config changes required.
### Contributors
- **danslapman** (#41, #42) — reported the Intel Mac build regression and investigated `ort-tract` as an alternative backend; closure documented that `ort-tract` returns `Unimplemented` when fastembed calls into it, confirming `ort-dynamic` as the correct path forward.
Hygiene release plus two UI gap closures. No breaking changes, no new major features, no schema migrations affecting user data beyond V11 dropping two verified-unused tables.
### Added
- **`POST /api/memories/{id}/suppress`** — Dashboard users can now trigger top-down inhibitory control (Anderson 2025 SIF + Davis Rac1 cascade) without dropping to raw MCP. Optional JSON body `{"reason": "..."}` logged for audit. Each call compounds; response includes `suppressionCount`, `priorCount`, `retrievalPenalty`, `reversibleUntil`, `estimatedCascadeNeighbors`, and `labileWindowHours`. Emits the existing `MemorySuppressed` WebSocket event so the 3D graph plays the violet implosion + compounding pulse shipped in v2.0.6.
- **`POST /api/memories/{id}/unsuppress`** — Reverses a suppression inside the 24h labile window. Returns `stillSuppressed: bool` so the UI can tell a full reversal from a compounded-down state. Emits `MemoryUnsuppressed` for the rainbow-burst reversal animation.
- **Suppress button on the Memories page** — Third action alongside Promote / Demote / Delete, hover-tooltip explaining the neuroscience ("Top-down inhibition (Anderson 2025). Compounds. Reversible for 24h.").
- **Uptime in the sidebar footer** — The Heartbeat WebSocket event has carried `uptime_secs` since v2.0.5 but was never rendered. Now displays as `up 3d 4h` / `up 18m` / `up 47s` (compact two-most-significant-units format) next to memory count + retention.
### Fixed
- **`execute_export` no longer panics on unknown format.** The write-out match arm at `maintenance.rs` was `_ => unreachable!()` — defensive `Err(...)` now returns a clean "unsupported export format" message instead of unwinding through the MCP dispatcher.
- **Dashboard graph page distinguishes empty-database from API failure** (landed in the first half of this branch). Before v2.0.7 any error from `/api/graph` rendered as "No memories yet," which masked real failures. Now the regex + node-count gate splits the two; real errors surface as "Failed to load graph: [sanitized message]" with filesystem paths stripped for info-disclosure hardening.
- **`predict` MCP tool surfaces a `predict_degraded` flag** instead of silently returning empty vecs on lock poisoning. `tracing::warn!` logs the per-channel error for observability.
- **`memory_changelog` honors `start` / `end` ISO-8601 bounds.** Previously advertised in the schema since v1.7 but runtime-ignored. Malformed timestamps now return a helpful error instead of silently dropping the filter. Response includes a `filter` field echoing the applied window.
- **`intention` check honors `include_snoozed`.** Previously silent no-op; snoozed intentions were invisible to check regardless of the arg. Dedup via HashSet guards against storage overlap.
- **`intention` check response exposes `status` and `snoozedUntil`** so callers can distinguish active-triggered from snoozed-overdue intentions.
- **Server tool-count comment at `server.rs:212`** updated (23 → 24) to match the runtime assertion.
### Removed
- **Migration V11: drops dead `knowledge_edges` + `compressed_memories` tables.** Both were added speculatively in V4 and marked deprecated in the same migration that created them. Zero INSERT or SELECT anywhere in `crates/`. Frees schema space for future migrations.
- **`execute_health_check` (71 LOC) + `execute_stats` (179 LOC) in `maintenance.rs`.** Both `#[allow(dead_code)]` since v1.7 with in-file comments routing users to `execute_system_status` instead. Zero callers workspace-wide. Net -273 LOC in the touched file.
- **`x86_64-apple-darwin` job from `.github/workflows/release.yml`.** The Intel Mac build failed the v2.0.5 AND v2.0.6 release workflows because `ort-sys 2.0.0-rc.11` (pinned by `fastembed 5.13.2`) does not ship Intel Mac prebuilts. `ci.yml` had already dropped the target; `release.yml` is now in sync. README documents the build-from-source path. Future releases should publish clean on all three supported platforms (macOS ARM64, Linux x86_64, Windows MSVC).
- Historical CHANGELOG entries and `docs/launch/*.md` launch materials left unchanged because they are time-stamped artifacts of their respective releases.
### Tests
- **+7 assertions** covering the v2.0.7 behavioral changes: V11 migration drops dead tables + is idempotent on replay, `predict_degraded` false on happy path, `include_snoozed` both paths + `status` field exposure, malformed `start` returns helpful error + `filter` field echo.
- Full suite: **1,292 Rust passing / 0 failed** across `cargo test --workspace --release`. **171 dashboard tests passing.** Zero clippy warnings on `vestige-core` or `vestige-mcp` under `-D warnings`.
### Audit
Pre-merge audited by 4 parallel reviewers (security, code quality, end-to-end flow trace, external verification). Zero CRITICAL or HIGH findings. Two MEDIUM fixes landed in the branch: graph error-message path sanitization (strip `/path/to/*.{sqlite,rs,db,toml,lock}`, cap 200 chars) and `intention` response `status` field exposure.
Polish release aimed at new-user happiness. v2.0.5's cognitive stack was already shipping; v2.0.6 makes it *feel* alive in the dashboard and stays out of your way on the prompt side.
### Added
#### Dashboard visual feedback for six live events
-`minimal` (default) — 3-sentence MCP `instructions` string telling the client how to use Vestige and how to react to explicit feedback. Safe for every audience, every client, every use case.
-`full` — opt in to the composition mandate (Composing / Never-composed / Recommendation response shape + FSRS-trust blocking phrase). Useful for high-stakes decision workflows; misfires on trivial retrievals, which is why it is not the default.
Advertised in `vestige-mcp --help` alongside `VESTIGE_DASHBOARD_ENABLED`.
### Fixed
#### Dashboard intentions page
-`IntentionItem.priority` was typed as `string` but the API returns the numeric FSRS-style scale (1=low, 2=normal, 3=high, 4=critical). Every intention rendered as "normal priority" regardless of its real value. Now uses a `PRIORITY_LABELS` map keyed by the numeric scale.
-`trigger_value` was typed as a plain string but the API returns `trigger_data` as a JSON-encoded payload (e.g. `{"type":"time","at":"..."}`). The UI surfaced raw JSON for every non-manual trigger. A new `summarizeTrigger()` helper parses `trigger_data` and picks the most human-readable field — `condition` / `topic` / formatted `at` / `in_minutes` / `codebase/filePattern` — before truncating for display. Closes the loop on PR #26's snake_case TriggerSpec fix at the UI layer.
### Docs
-`README.md` — new "What's New in v2.0.6" header up top; v2.0.5 block strengthened with explicit contrast against Ebbinghaus 1885 passive decay and Anderson 1994 retrieval-induced forgetting; new "Forgetting" row in the RAG-vs-Vestige comparison table.
- Intel-Mac and Windows install steps replaced with a working `cargo build --release -p vestige-mcp` snippet. The pre-built binaries for those targets are blocked on upstream toolchain gaps (`ort-sys` lacks Intel-Mac prebuilts in the 2.0.0-rc.11 release pinned by `fastembed 5.13.2`; `usearch 2.24.0` hit a Windows MSVC compile break tracked as [usearch#746](https://github.com/unum-cloud/usearch/issues/746)).
### Safety
No regressions of merged contributor PRs — v2.0.6 only touches regions that are non-overlapping with #20 (resource URI strip), #24 (codex integration docs), #26 (snake_case TriggerSpec), #28 (deep_reference query relevance), #29 (older glibc feature flags), #30 (`VESTIGE_DASHBOARD_ENABLED`), #32 (dream eviction), and #33 (keyword-first search).
Every AI memory system stores too much. Vestige now treats forgetting as a first-class, neuroscientifically-grounded primitive. This release adds **active forgetting** — top-down inhibitory control over memory retrieval, based on two 2025 papers that no other AI memory system has implemented.
### Scientific grounding
- **Anderson, M. C., Hanslmayr, S., & Quaegebeur, L. (2025).** *"Brain mechanisms underlying the inhibitory control of thought."* Nature Reviews Neuroscience. DOI: [10.1038/s41583-025-00929-y](https://www.nature.com/articles/s41583-025-00929-y). Establishes the right lateral PFC as the domain-general inhibitory controller, and Suppression-Induced Forgetting (SIF) as compounding with each stopping attempt.
- **Cervantes-Sandoval, I., Chakraborty, M., MacMullen, C., & Davis, R. L. (2020).** *"Rac1 Impairs Forgetting-Induced Cellular Plasticity in Mushroom Body Output Neurons."* Front Cell Neurosci. [PMC7477079](https://pmc.ncbi.nlm.nih.gov/articles/PMC7477079/). Establishes Rac1 GTPase as the active synaptic destabilization mechanism — forgetting is a biological PROCESS, not passive decay.
### Added
#### `suppress` MCP Tool (NEW — Tool #24)
- **Top-down memory suppression.** Distinct from `memory.delete` (which removes) and `memory.demote` (which is a one-shot hit). Each `suppress` call compounds: `suppression_count` increments, and a `k × suppression_count` penalty (saturating at 80%) is subtracted from retrieval scores during hybrid search.
- **Rac1 cascade.** Background worker piggybacks the existing consolidation loop, walks `memory_connections` edges from recently-suppressed seeds, and applies attenuated FSRS decay to co-activated neighbors. You don't just forget "Jake" — you fade the café, the roommate, the birthday.
- **Reversible 24h labile window** — matches Nader reconsolidation semantics on a 24-hour axis. Pass `reverse: true` within 24h to undo. After that, it locks in.
- **Never deletes** — the memory persists and is still accessible via `memory.get(id)`. It's INHIBITED, not erased.
#### `active_forgetting` Cognitive Module (NEW — #30)
-`crates/vestige-core/src/neuroscience/active_forgetting.rs` — stateless helper for SIF penalty computation, labile window tracking, and Rac1 cascade factors.
- 7 unit tests + 9 integration tests = 16 new tests.
- Partial indices on both columns for efficient sweep queries.
- Additive-only — backward compatible with all existing v2.0.x databases.
#### Dashboard
-`ForgettingIndicator.svelte` — new status pill that pulses when suppressed memories exist.
- 3D graph nodes dim to 20% opacity and lose emissive glow when suppressed.
- New WebSocket events: `MemorySuppressed`, `MemoryUnsuppressed`, `Rac1CascadeSwept`.
-`Heartbeat` event now carries `suppressed_count` for live dashboard display.
### Changed
-`search` scoring pipeline now includes an SIF penalty applied after the accessibility filter.
- Consolidation worker (`VESTIGE_CONSOLIDATION_INTERVAL_HOURS`, default 6h) now runs `run_rac1_cascade_sweep` after each `run_consolidation` call.
- Tool count assertion bumped from 23 → 24.
- Workspace version bumped 2.0.4 → 2.0.5.
### Tests
- Rust: 1,284 passing (up from 1,237). Net +47 new tests for active forgetting, Rac1 cascade, migration V10.
- Dashboard (Vitest): 171 passing (up from 150). +21 regression tests locking in the issue #31 UI fix.
- Zero warnings, clippy clean across all targets.
### Fixed
- **Dashboard graph view rendered glowing squares instead of round halos** ([#31](https://github.com/samvallad33/vestige/issues/31)). Root cause: the node glow `THREE.SpriteMaterial` had no `map` set, so `Sprite` rendered as a solid-coloured 1×1 plane; additive blending plus `UnrealBloomPass(strength=0.8, radius=0.4, threshold=0.85)` then amplified the square edges into hard-edged glowing cubes. The aggressive `FogExp2(..., 0.008)` swallowed edges at depth and dark-navy `0x4a4a7a` lines were invisible against the fog. Fix bundled:
- Generated a shared 128×128 radial-gradient `CanvasTexture` (module-level singleton) and assigned it as `SpriteMaterial.map`. Gradient stops: `rgba(255,255,255,1.0) → rgba(255,255,255,0.7) → rgba(255,255,255,0.2) → rgba(255,255,255,0.0)`. Sprite now reads as a soft round halo; bloom diffuses cleanly.
- Retuned `UnrealBloomPass` to `(strength=0.55, radius=0.6, threshold=0.2)` — gentler, allows mid-tones to bloom instead of only blown-out highlights.
- Halved fog density `FogExp2(0x050510, 0.008) → FogExp2(0x0a0a1a, 0.0035)` so distant memories stay visible.
- Bumped edge color `0x4a4a7a → 0x8b5cf6` (brand violet). Opacity `0.1 + weight*0.5 → 0.25 + weight*0.5`, cap `0.6 → 0.8`. Added `depthWrite: false` so edges blend cleanly through fog.
- Added explicit `scene.background = 0x05050f` and a 2000-point starfield distributed on a spherical shell at radius 600–1000, additive-blended with subtle cool-white/violet vertex colors.
- Glow sprite scale bumped `size × 4 → size × 6` so the gradient has visible screen footprint.
- All node glow sprites share a single `CanvasTexture` instance (singleton cache — memory leak guard for large graphs).
- 21 regression tests added in `apps/dashboard/src/lib/graph/__tests__/ui-fixes.test.ts`. Hybrid strategy: runtime unit tests via the existing `three-mock.ts` (extended to propagate `map`/`color`/`depthWrite`/`blending` params and added `createRadialGradient` to the canvas context mock), plus source-level regex assertions on `scene.ts` and `nodes.ts` magic numbers so any accidental revert of fog/bloom/color/helper fails the suite immediately.
-`apps/dashboard/package.json` version stale at 2.0.3 — bumped to 2.0.5 to match the workspace.
-`packages/vestige-mcp-npm/.gitignore` missing `bin/vestige-restore` and `bin/vestige-restore.exe` entries — the other three binaries were already ignored as postinstall downloads.
Context windows hit 1M tokens. Memory matters more than ever. This release removes artificial limits, adds contradiction detection, and hardens security.
### Added
#### cross_reference Tool (NEW — Tool #22)
- **Connect the dots across memories.** Given a query or claim, searches broadly, detects agreements and contradictions between memories, identifies superseded/outdated information, and returns a confidence-scored synthesis.
- Pairwise contradiction detection using negation pairs + correction signals, gated on shared topic words to prevent false positives.
- **Real-time 3D animations** — memories pulse on access, burst particles on creation, shockwave rings on dreams, golden flash lines on connection discovery, fade on decay
- **Bloom post-processing** — cinematic neural network aesthetic with UnrealBloomPass
- **GPU instanced rendering** — 1000+ nodes at 60fps via Three.js InstancedMesh
- **`token_budget` parameter on `search`** — limits response size (100-10000 tokens). Results exceeding budget moved to `expandable` array with `tokensUsed`/`tokenBudget` tracking.
- **Reader/writer connection split** — `Storage` struct uses `Mutex<Connection>` for separate reader/writer SQLite handles with WAL mode. All methods take `&self` (interior mutability). `Arc<Mutex<Storage>>` → `Arc<Storage>` across ~30 files.
- **Tool consolidation: 23 → 18 tools** — merged redundant tools while maintaining 100% backward compatibility via deprecated redirects
- **`ingest` → `smart_ingest`** — `ingest` was a duplicate of `smart_ingest`; now redirects automatically
- **`session_checkpoint` → `smart_ingest` batch mode** — new `items` parameter on `smart_ingest` accepts up to 20 items, each running the full cognitive pipeline (importance scoring, intent detection, synaptic tagging, hippocampal indexing). Old `session_checkpoint` skipped the cognitive pipeline.
- **`promote_memory` + `demote_memory` → `memory` unified** — new `promote` and `demote` actions on the `memory` tool with optional `reason` parameter and full cognitive feedback pipeline (reward signal, reconsolidation, competition)
- **`health_check` + `stats` → `system_status`** — single tool returns combined health status, full statistics, FSRS preview, cognitive module health, state distribution, warnings, and recommendations
- **CLAUDE.md automation overhaul** — all 18 tools now have explicit auto-trigger rules; session start expanded to 5 steps (added `system_status` + `predict`); full proactive behaviors table
### Added
-`smart_ingest` batch mode with `items` parameter (max 20 items, full cognitive pipeline per item)
-`memory` actions: `promote` and `demote` with optional `reason` parameter
-`system_status` tool combining health check + statistics + cognitive health
- 30 new tests (305 → 335)
### Deprecated (still work via redirects)
-`ingest` → use `smart_ingest`
-`session_checkpoint` → use `smart_ingest` with `items`
-`promote_memory` → use `memory(action="promote")`
