v2.1.2 Honest Memory

Concrete search, irreversible purge, first-class contradictions tool, vestige update CLI, dense dream persistence fix, embedding-model upgrade repair, and a /dashboard/waitlist Pro early-access preview.

25 MCP tools. SQLite migration v13. Backwards compatible: 'delete' remains as a 'purge' alias.

Closes #50, #51.

docs/COGNITIVE_SANDWICH.md

@@ -55,7 +55,7 @@ For each claim, it checks Vestige's `deep_reference` for high-trust contradictin
 |---|---|
 | TECHNICAL / EXISTENTIAL / TIMELINE | VETO if memory trust > 0.55 directly contradicts |
 | BIOGRAPHICAL / FINANCIAL / ACHIEVEMENT / ATTRIBUTION | VETO if contradicted OR if factual-shaped with zero supporting evidence (fail-closed) |
-| **VAGUE-QUANTIFIER** | VETO on "a few wins / some prize money / most placed" without enumeration |
+| **VAGUE-QUANTIFIER** | VETO on vague achievement or financial claims without enumeration |
 | **UNVERIFIED-POSITIVE** | VETO on specific named institutions/dates/employers not in evidence |
 
 False-positive guards (added v2.1.0 after dogfood):
@@ -183,7 +183,7 @@ On M3 Max 14-core or M2/M1 Max: closer to 3–7s prompt processing, ~50–60 tok
 
 ## Architecture provenance
 
-The Cognitive Sandwich originated April 2026 as a defense against the AIMO3 36/50 failure mode — Claude retrieving relevant memories but summarizing them instead of composing them into recommendations. The pre-cognitive layer enforces composition; the post-cognitive layer catches contradictions before they ship.
+The Cognitive Sandwich originated April 2026 as a defense against a dogfood failure mode: Claude retrieved relevant memories but summarized them instead of composing them into a recommendation. The pre-cognitive layer enforces composition; the post-cognitive layer catches contradictions before they ship.
 
 Full architecture memory: search Vestige for `god-tier-plan` or `cognitive-sandwich` tags after install.
 

docs/FAQ.md

@@ -160,6 +160,8 @@ Accessibility is calculated as: `0.5 × retention + 0.3 × retrieval_strength +
 
 Memories are never deleted automatically. They fade from relevance but can be revived if accessed again (like human memory—"oh, I forgot about that!").
 
+If you explicitly want content gone, use `memory(action="purge", confirm=true)`. Purge permanently removes the memory content and embeddings, scrubs internal references, and keeps only a non-content tombstone so sync/audit can prove the deletion happened.
+
 **To configure decay**: The FSRS-6 algorithm auto-tunes based on your usage patterns. Memories you access stay strong; memories you ignore fade. No manual tuning needed.
 </details>
 

docs/STORAGE.md

@@ -54,7 +54,7 @@ vestige portable-import ~/Dropbox/vestige/portable.json --merge
 vestige sync ~/Dropbox/vestige/portable.json
 ```
 
-`vestige sync` uses the same pluggable portable-sync backend interface as the core library. v2.1.1 ships a file backend, which works with Dropbox, iCloud Drive, Syncthing, Git, network shares, or any folder-sync system. The merge algorithm applies delete tombstones, keeps newer local memories on timestamp conflicts, preserves stable IDs, rebuilds FTS after import, and writes the pushed archive atomically when the filesystem supports rename.
+`vestige sync` uses the same pluggable portable-sync backend interface as the core library. v2.1.1 ships a file backend, which works with Dropbox, iCloud Drive, Syncthing, Git, network shares, or any folder-sync system. The merge algorithm applies delete tombstones, keeps newer local memories on timestamp conflicts, preserves stable IDs, rebuilds FTS after import, and writes the pushed archive atomically when the filesystem supports rename. v2.1.2 also carries non-content purge tombstones so a hard purge can sync without retaining the deleted memory text.
 
 When using the MCP `export` tool with `format: "portable"`, Vestige writes the archive under the active data directory's `exports/` folder. The MCP `restore` tool only reads from that `exports/` or `backups/` folder by default; pass `allowAnyPath: true` only for a trusted local file you selected manually.
 

docs/integrations/windsurf.md

@@ -115,7 +115,7 @@ It remembers.
 
 ## Important: Tool Limit
 
-Windsurf has a **hard cap of 100 tools** across all MCP servers. Vestige uses 24 tools, leaving plenty of room for other servers.
+Windsurf has a **hard cap of 100 tools** across all MCP servers. Vestige uses 25 tools, leaving plenty of room for other servers.
 
 ---
 

docs/integrations/xcode.md

@@ -50,7 +50,7 @@ Quit Xcode completely (Cmd+Q) and reopen your project.
 
 ### 4. Verify
 
-Type `/context` in the Agent panel. You should see `vestige` listed with 24 tools.
+Type `/context` in the Agent panel. You should see `vestige` listed with 25 tools.
 
 ---
 

docs/launch/UI_ROADMAP_v2.1_v2.2.md

@@ -1,201 +0,0 @@
-# Vestige UI Roadmap — v2.1.0 and v2.2.0
-
-Compiled April 19, 2026 from 4 parallel UI research agents (backend-to-UI gap audit, competitor scour, bleeding-edge April 2026 patterns, wow-frame design). Local-only planning doc — not for commit to main until scope is locked.
-
----
-
-## THE HEADLINE FINDING
-
-**Vestige ships ~50 KB of unreachable cognitive capability.** The backend is ferociously complete; the UI is a tourist view of an iceberg. Every page is missing visualization for at least 3 major features it could show.
-
-- **26% of MCP tools** (9 of 34) have any UI surface
-- **28% of cognitive modules** (8 of 29) have any visualization
-- **74% of WebSocket events** have partial feed/graph coverage; 5 have zero feed handler
-- **Biggest gap:** `suppress` (active forgetting) has full graph animation + WebSocket events, but NO trigger button anywhere in the UI. Users literally cannot trigger the signature v2.0.5 feature from the dashboard.
-
-The v2.1.0 UI story writes itself: **"Vestige v2.1 makes the invisible visible."**
-
----
-
-## TOP 10 CRITICAL UI GAPS (from Agent 1, ordered by user-visible impact)
-
-1. **`suppress` tool has zero frontend trigger.** Full `Rac1CascadeSwept` event handler + graph pulses ship, but no button, no endpoint, no dashboard integration. Users can't forget anything without raw MCP access.
-2. **Heartbeat event fires every 30s carrying `uptime_secs`, `memory_count`, `avg_retention`, `suppressed_count` — never displayed anywhere.** Real-time health that costs nothing to show.
-3. **`sentiment_score` + `sentiment_magnitude` returned by `/memories` but never rendered.** Emotional coloring is invisible.
-4. **Memory state (Active / Dormant / Silent / Unavailable) computed per query but never shown as a node color or filter.**
-5. **Intention page is list-only.** No endpoints for status change, snooze, or complete. Users can see intentions but not act on them from the dashboard.
-6. **Rac1 cascade shows animation with zero data summary.** Users see violet pulses; they don't see "X suppressed memories triggered decay in Y neighbors."
-7. **Synaptic tagging 9h window is invisible.** Retroactive importance boost happens silently.
-8. **Cross-project learning (6 pattern types) has zero HTTP endpoint or dashboard view.**
-9. **Consolidation internals hidden.** Which nodes decayed, which got new embeddings — all computed, all hidden.
-10. **`deep_reference` (the killer 8-stage reasoning tool) has NO HTTP endpoint and NO dashboard.** The v2.0.4 headline feature is unreachable from the UI.
-
----
-
-## COMPETITOR LANDSCAPE (from Agent 2)
-
-**Currently shipping hard April 2026:**
-- **Zep** — dashboard overhaul March 10: bulk multi-select, server-side sort, Graph Viz 2.0 (nodes sized by connection count, no render cap, click-node details). Closest competitor on graph.
-- **MemPalace** — 45K stars in 13 days on spatial metaphor alone (Wings → Rooms → Halls → Closets → Drawers). 13 releases in 13 days.
-- **Cognee v0.3.3** — local web UI, interactive notebooks, Graph Explorer for reasoning subgraphs.
-- **Letta ADE** — 3-panel Agent Development Environment at app.letta.com. Context window viewer, memory blocks, archival search.
-
-**Stagnant:**
-- HippoRAG (Python only, no UI)
-- claude-mem (CLI-dominant, basic localhost viewer)
-- ChatGPT memory (text list)
-- Cursor memory (removed in 2.1)
-
-**What NOBODY has (unclaimed UI territory):**
-1. Ambient always-on memory widget (menu bar / tray)
-2. Watch / ring interface
-3. Voice-first memory UI
-4. Collaborative multi-user graph (Figma cursors for memory)
-5. AR/VR memory palace (native Vision Pro / Quest)
-6. Temporal time-scrubber (drag slider to rewind graph state)
-7. Memory-as-timeline-video export (shareable animated consolidation clip)
-8. Contradiction surfacing UI ("Disputes" page)
-9. FSRS retention heatmap calendar (GitHub-contribution-grid style)
-10. Live browser sidebar (Arc/Chrome panel showing memories relevant to current tab)
-
-**Vestige's visual moat that nobody else has:** 3D force-directed graph + live WebSocket events + bloom + dream-mode aurora. Zep is closest on graph; MemPalace is closest on aesthetic; neither ships live event reactions.
-
----
-
-## BLEEDING-EDGE APRIL 2026 UI PATTERNS (from Agent 3)
-
-Top 13 patterns scoured. The 5 most applicable to Vestige:
-
-1. **Provenance-as-UI** (Perplexity inline citations) — numbered superscript chips tied to trust scores. Vestige has FSRS trust; just doesn't surface it inline.
-2. **Ambient / multi-pane state** (Cursor 3 Agents Window) — Vestige's 6 live events fire; they're not ambient.
-3. **Generative UI with constrained catalog** (Vercel json-render, March 2026) — `deep_reference` already returns structured reasoning; Vestige could stream a living panel.
-4. **Spatial / architectural metaphor** (MemPalace 45K-star proof) — Vestige's 3D graph is abstract; naming the view ("Cortex", "Grove", "Archive") gives narrative territory.
-5. **Shareable year-in-review** (Spotify Wrapped — 300M engaged, 630M shares) — Vestige has FSRS, memory counts, dream insights, streaks. All the ingredients for a free distribution loop.
-
-**Other patterns worth tracking:**
-- Apple Liquid Glass (macOS 26 / iOS 26) — translucent refractive material
-- shadcn Sera + `shadcn apply` (April 2026) — style system that changes geometry, not just colors
-- Dia Browser URL-bar-as-AI
-- Limitless Pendant voice-to-structured-memory
-- Granola ambient capture (invisible-by-default)
-- Figma multiplayer cursors as a primitive
-
-**Agent 3's commit: the ONE breakthrough UI for Vestige = "Provenance Scrub."**
-
-Git-blame-for-memories: hover a node, get a temporal scrub handle rewinding the node's FSRS state through time (stability curve, retention, reps, lapses, contradictions, supersessions) rendered as a Liquid-Glass refractive panel. Click any point on the scrub to see memory content at that time. Inline Perplexity citations tag every fact.
-
-Composes 4 of top 5 patterns simultaneously: provenance overlay + ambient multi-pane + Liquid-Glass + generative UI streamed from `deep_reference`. Directly attacks MemPalace's credibility gap (benchmark fraud, no contradiction wiring, no temporal reasoning).
-
-Engineering cost: 9 days. Floor: 3D scrub + trust chips in 4 days as v2.1 patch. Ceiling: full Liquid-Glass + generative panel as v2.2 headlining launch.
-
----
-
-## WOW FRAMES (from Agent 4) — ranked by ship priority
-
-### Ship in v2.1.0 (5.5 engineering days, two HN thumbnails)
-
-**1. Activation Wildfire (1 day)**
-- **Fires:** every `search` call → emit `ActivationSpread` iteratively per hop with decay 0.7.
-- **Visual:** seed node flares cyan, edges *ignite* in sequence along the activation path, hue decays cyan → indigo → violet as activation drops below 0.1.
-- **Neuroscience:** Collins & Loftus 1975, `spreading_activation.rs:1-58`.
-- **Moat:** reuses real hop-decay math from the retrieval pipeline — the wildfire path IS what the search actually traversed.
-
-**2. Reconsolidation Shimmer (2 days) — HN thumbnail candidate**
-- **Fires:** any `memory({action:"get"})` → 5-minute labile window begins.
-- **Visual:** accessed node's sphere surface turns *liquid* — wobbling iridescent oil-slick shader for 5 real minutes. Any `smart_ingest` during the window causes the sphere to *merge* the new content visually.
-- **Neuroscience:** Nader 2000, `reconsolidation.rs:405`.
-- **Moat:** a memory being *editable only when recalled* is pure Nader. The shimmer is the meme shot.
-
-**3. Dream Stitching (2.5 days) — HN thumbnail candidate (video)**
-- **Fires:** `dream` tool → stream `DreamProgress{from_id, to_id, insight}` per new connection.
-- **Visual:** camera auto-orbits into existing dream-mode aurora. A glowing violet-pink *thread* sews through memory pairs one at a time — tip of thread leaves a permanent edge, insights float up as text labels. Ends with a supernova at graph centroid.
-- **Neuroscience:** MemoryDreamer 5-stage consolidation.
-- **Moat:** dreams *creating new edges* is Vestige-exclusive.
-
-### Queue for v2.2.0
-
-**4. Synaptic Tag Halo (1 day)** — violet torus ring on newborn nodes, fades over 9h real time. Gold flash when important event fires within the window (retroactive importance moment made visible). `synaptic_tagging.rs`.
-
-**5. Competition Duel (1 day)** — top-3 search results duel. Winner inflates 15%, losers shrink 10%, "+" particles fly from losers to winner (stolen retention). Anderson 1994 retrieval-induced forgetting.
-
-**6. Rac1 Slow Burn (1.5 days)** — suppressed seed blackens into graphite. Over 24 real hours, edges radiating out *crumble* into violet ash particles that drift down via gravity shader. Dead branches literally fall away.
-
-**7. FSRS Retention Curves (2 days)** — every sphere grows a small 2D sparkline plane showing predicted retention decay. Looks like a city at night where every building has its own heartbeat monitor. Nodes approaching Dormant threshold pulse amber.
-
----
-
-## COMPOSED v2.1.0 AND v2.2.0 UI ROADMAP
-
-### v2.1.0 "Decide" (May 5-6 launch) — UI track
-
-On top of the already-planned v2.1.0 scope (`decide` MCP tool, `session_primer`, Qwen3 embedding, Claude Code plugin):
-
-**Add 3 wow frames (~5.5 days):**
-1. Activation Wildfire — 1 day
-2. Reconsolidation Shimmer — 2 days (HN thumbnail screenshot)
-3. Dream Stitching — 2.5 days (HN thumbnail video)
-
-**Add 5 of the top-10 gap fixes (~5 days):**
-1. `suppress` trigger button + HTTP endpoint — 1 day
-2. Heartbeat display widget (uptime + avg retention + suppressed count) — 0.5 day
-3. Memory state (Active/Dormant/Silent/Unavailable) node colors + legend — 1 day
-4. Intention update/snooze/complete endpoints + UI — 1 day
-5. `deep_reference` dashboard page (the 8-stage reasoning viewer) — 1.5 days
-
-**Total v2.1.0 UI scope: ~10.5 engineering days** on top of the existing 19.5 day Qwen3 + decide + plugin scope. Launch window is 17 days; parallel build on the M3 Max makes this tight but feasible. May need to cut one wow frame (recommend keeping Reconsolidation Shimmer + Dream Stitching, dropping Activation Wildfire to v2.1.1 if time-pressed).
-
-### v2.2.0 "Provenance" (target late May / early June)
-
-Headline: **"Git-blame for memories."** The Provenance Scrub compose (Agent 3's breakthrough).
-
-- 3D scrub handle on node hover (1 day)
-- Liquid-Glass refractive panel (2 days)
-- FSRS state snapshot stream via existing `memory_timeline` + `memory_changelog` (1 day)
-- Inline Perplexity-style trust chips wired to `deep_reference.evidence[]` (1.5 days)
-- Generative side-panel streaming `deep_reference.reasoning` json-render-style (2 days)
-- Polish + demo clip (1.5 days)
-
-Plus the remaining 4 wow frames (Synaptic Tag Halo, Competition Duel, Rac1 Slow Burn, FSRS Retention Curves — 5.5 days).
-
-**Total v2.2.0 UI scope: ~14.5 days.** Ship target: June graduation week (June 13).
-
-### v2.3.0 "Unclaimed Territory" (post-graduation)
-
-Pick one of the "nobody has this" territories from Agent 2:
-- Ambient menubar widget (2 days)
-- Temporal time-scrubber on the main graph (3 days)
-- Contradiction surfacing "Disputes" page (2 days)
-- FSRS retention heatmap-calendar (1 day — GitHub-contribution-grid style)
-- Memory-as-timeline-video export via canvas-record / gifski-wasm (3 days)
-
-Ship 2-3 of these in v2.3. Each is an unclaimed moat.
-
----
-
-## WHAT NOT TO DO
-
-- **Don't add memory palace metaphor (Wings/Rooms/Halls).** MemPalace owns that narrative territory with 45K stars. Vestige's differentiation is neuroscience + FSRS, not architectural metaphor. Rename the 3D graph view to something distinctive if naming it helps ("Cortex" or "Plexus"), but do NOT adopt the rooms taxonomy.
-- **Don't chase every 2026 pattern.** Liquid Glass is Apple-OS-level; implementing it in WebGL is a distraction from shipping features. Save for v2.2 selectively.
-- **Don't build mobile yet.** Adoption curve isn't there. Desktop dashboard + MCP server first.
-- **Don't build multi-user.** Single-user local is the AGPL-3.0 story. Multi-tenant is vestige-cloud (proprietary), separate roadmap.
-
----
-
-## Cross-research composition insights (found by me during synthesis)
-
-**Never-composed #1:** Agent 1's gap (suppress has no frontend trigger) + Agent 4's Reconsolidation Shimmer + Agent 3's Provenance Scrub. Three pieces of the "make the invisible visible" story. Ship them together as v2.1.0 UI narrative.
-
-**Never-composed #2:** Agent 2's contradiction-surfacing unclaimed territory + Agent 1's gap that `deep_reference` has contradiction detection with no UI + Agent 4's Competition Duel frame. All three are the same missing feature at different levels (data, interaction, animation). Ship as v2.2 "Disputes" page + Competition Duel micro-animation together.
-
-**Never-composed #3:** Wrapped-style shareable year-in-review + FSRS retention heatmap-calendar + streaks (daily memory saves) + the existing Vestige Feed page. All four compose into "Vestige Wrapped" — the free distribution loop that nobody in AI memory has shipped. Ship as v2.3 "Year in Memory" — summer 2026, after launch stabilizes.
-
----
-
-## What this document is FOR
-
-- **Reference** when scoping v2.1.0 and v2.2.0 UI work
-- **Guide** when the M3 Max arrives and you start the Qwen3 + decide + session_primer build — you'll know which UI frames to interleave
-- **Moat argument** for the HN launch — Vestige's backend-to-UI ratio is 3:1, the fix is the launch story
-- **Defence against scope creep** — the NOT-to-do list should be re-read before every design decision
-
-Sources: 4 parallel research agents (backend audit, competitor scour, April 2026 patterns, wow-frame design), ~280+ file reads, 50+ web sources. Full raw outputs preserved in Claude Code session logs.

docs/launch/demo-script.md

@@ -419,7 +419,7 @@ vestige-mcp --version
 > Yes. It speaks MCP — the Model Context Protocol. One config change and it works with Claude Desktop, Cursor, VS Code Copilot, JetBrains, Windsurf, Xcode 26.3. Anything that speaks MCP.
 
 **Q: What about multi-user or team memory?**
-> That's the v3.0 roadmap — "Hivemind." Ed25519 identity, CRDT-based sync, transactive directory (Wegner's "who knows what" routing), federated retrieval with differential privacy. The open source version is single-user, local-first. Team and cloud features will be proprietary.
+> The current Pro plan is more pragmatic: prove portable sync/storage first, then ship Solo and Team workflows around managed sync, backups, onboarding, and support. The open-source core stays local-first; paid team features should stay in the separate Pro/commercial boundary.
 
 **Q: How does Prediction Error Gating prevent duplicate memories?**
 > When you ingest a new memory, it computes embedding similarity against all existing memories. If similarity is above 0.92, it reinforces the existing memory (bumps FSRS stability). Between 0.75 and 0.92, it updates/merges. Below 0.75, it creates a new memory. The thresholds come from computational neuroscience research on prediction error signals — the brain stores what's surprising, reinforces what's familiar, and updates what's partially known. Same principle.

docs/launch/reddit-cross-reference.md

@@ -10,11 +10,11 @@ I've been building Vestige — an MCP memory server that gives Claude persistent
 
 But last week it almost cost me hours of debugging.
 
-Claude confidently told me my AIMO3 competition notebook should use `--enable-prefix-caching` with vLLM. I trusted it. The notebook crashed. Scored 0/50. Burned a daily submission.
+Claude confidently told me a benchmark notebook should use `--enable-prefix-caching` with vLLM. I trusted it. The notebook crashed. Burned a daily submission.
 
 The problem? I had TWO memories:
 - **January**: "prefix caching crashes with our vLLM build"  
-- **March**: "prefix caching works with the new animsamuelk wheels"
+- **March**: "prefix caching works with the newer vLLM build"
 
 Claude found both. Picked the wrong one. Gave me a confident wrong answer based on the January memory. The March memory was correct — but Claude had no way to know they conflicted.
 
@@ -38,11 +38,11 @@ And gets back:
     {
       "newer": {
         "date": "2026-03-18",
-        "preview": "Switched to animsamuelk wheels which support --enable-prefix-caching..."
+        "preview": "Switched to a newer vLLM build that supports --enable-prefix-caching..."
       },
       "older": {
         "date": "2026-01-15", 
-        "preview": "prefix caching crashed with our samvalladares vLLM build..."
+        "preview": "prefix caching crashed with our older local vLLM build..."
       },
       "recommendation": "Trust the newer memory. Consider demoting the older one."
     }