apunkt/vestige
1
0
Fork 0
mirror of https://github.com/samvallad33/vestige.git synced 2026-06-08 20:25:16 +02:00
Code Issues Projects Releases Packages Wiki Activity
vestige/CHANGELOG.md
Sam Valladares 4e9e11ac0b Fix dense dream connection persistence 
Fixes #50
2026-05-01 05:37:27 -05:00

51 KiB
Raw Blame History

Changelog

All notable changes to Vestige will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Fixed

  • 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.

[2.1.1] - 2026-05-01 — "Portable Sync"

v2.1.1 focuses on user-controlled portability: exact storage archives, merge-safe file sync, pluggable sync backends, and explicit hook opt-ins.

Added

  • Exact portable archives — vestige portable-export / vestige portable-import preserve raw Vestige storage rows: memory IDs, FSRS state, graph edges, suppression state, audit rows, sessions, intentions, and embedding blobs.
  • 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.

[2.1.0] - 2026-04-27 — "Cognitive Sandwich Goes Local"

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.

Added

  • 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).
    • settings.fragment.json — empty default JSON snippet; the installer only wires hooks from the opt-in preflight/Sanhedrin fragments.
  • 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.
  • agents/ — executioner.md (legacy/fallback Haiku 4.5 path), lateral-thinker.md, synthesis-composer.md.
  • 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.
  • docs/COGNITIVE_SANDWICH.md — architecture diagram, install guide, performance notes (82 tok/s on M3 Max), uninstall, configuration env vars.
  • PR #48 — VESTIGE_DATA_DIR env-var support + tilde expansion + secure unix perms (thanks @Jelloeater) — directly addresses the ghost env-vars exposed by v2.0.9 cleanup.

Changed

  • 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.

Verified

  • cargo test --workspace --release --no-fail-fast: 1,229 passing, 0 failed (366 vestige-core + 358 vestige-mcp lib + 4 vestige-mcp bin + 497 e2e + 4 doctests).
  • Sanhedrin bridge smoke checks: Python bytecode compilation passes, fail-open bridge invocation returns yes, and public hook settings validate as JSON.
  • v2.1.0 hotfix installer matrix: default, preflight-only, Sanhedrin-only, full sandwich, legacy-all-hooks migration, and unrelated custom hooks preservation.
  • 8-day Sandwich dogfood: 84% pass rate, 16% legitimate vetoes caught real hallucinations.

Closes

  • #36 (Agent Hooks for Low-Effort Automatic Memory Capture) — Cognitive Sandwich is the answer.

Prerequisites for optional local MLX Sanhedrin

  • macOS Apple Silicon (M1+) — required only for --with-launchd MLX autostart
  • Python 3.10+
  • ~22 GB free RAM (Qwen3.6-35B-A3B-4bit at runtime)
  • First-run model download: ~19 GB from Hugging Face (cached locally thereafter)

Migration

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.

[2.0.9] - 2026-04-24 — "Autopilot"

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:

    • MemoryCreated → synaptic_tagging.trigger_prp() (9h retroactive PRP window, Frey & Morris 1997) + predictive_memory.record_memory_access().
    • 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.
  • cargo clippy -p vestige-mcp --lib --bins -- -D warnings: clean.
  • Five-agent parallel audit (security, dead-code, flow-trace, runtime-safety, release-prep): all GO.

Migration

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.

[2.0.8] - 2026-04-23 — "Pulse"

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.
  • 7 additional dashboard surfaces exposing the cognitive engine (v2.4 UI expansion): /duplicates (find_duplicates cluster view), /dreams (5-stage replay + insight cards), /schedule (FSRS calendar + retention forecast), /importance (4-channel novelty/arousal/reward/attention radar), /activation (spreading-activation network viz), /contradictions (trust-weighted conflict arcs), /patterns (cross-project pattern-transfer heatmap). Left nav expanded from 8 → 16 entries with single-key shortcuts (R/A/D/C/P/U/X/N).
  • 3D Graph brightness system — auto distance-compensated node brightness (1.0× at camera <60u, up to 2.4× at far zoom) so nodes don't disappear into exponential fog at zoom-out. User-facing brightness slider in the graph toolbar (☀ icon, range 0.5×-2.5×, localStorage-persisted under vestige:graph:brightness). Composes with the auto boost; opacity + glow halo + edge weight track the combined multiplier so nodes stay coherent.
  • 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.

[2.0.7] - 2026-04-19 — "Visible"

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).

Docs

  • Reconciled tool / module / test counts across README.md, CONTRIBUTING.md, docs/integrations/windsurf.md, docs/integrations/xcode.md. Ground truth: 24 MCP tools · 29 cognitive modules · 1,292 Rust tests + 171 dashboard tests.
  • 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.

[2.0.6] - 2026-04-18 — "Composer"

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

  • MemorySuppressed → violet implosion + compounding pulse whose intensity scales with suppression_count (Anderson 2025 SIF visualised).
  • MemoryUnsuppressed → rainbow burst + green pulse when a memory is brought back within the 24h labile window.
  • Rac1CascadeSwept → violet wave across a random neighbour sample while the background Rac1 worker fades co-activated memories.
  • Connected → gentle cyan ripple on WebSocket handshake.
  • ConsolidationStarted → subtle amber pulses across a 20-node sample during the FSRS-6 decay cycle (matches feed-entry colour).
  • ImportanceScored → magenta pulse on the scored node with intensity proportional to composite score.

Before v2.0.6 all six events fired against a silent graph. Users perceived the dashboard as broken or unresponsive during real cognitive work.

VESTIGE_SYSTEM_PROMPT_MODE environment variable

  • 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).

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).

[2.0.5] - 2026-04-14 — "Intentional Amnesia"

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. 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. 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.

Migration V10

  • ALTER TABLE knowledge_nodes ADD COLUMN suppression_count INTEGER DEFAULT 0
  • ALTER TABLE knowledge_nodes ADD COLUMN suppressed_at TEXT
  • 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). 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.

[2.0.4] - 2026-04-09 — "Deep Reference"

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.
  • Timeline analysis (newest-first), confidence scoring (agreements boost, contradictions penalize, recency bonus).

retrieval_mode Parameter (search tool)

  • precise — top results only, no spreading activation or competition. Fast, token-efficient.
  • balanced — full 7-stage cognitive pipeline (default, no behavior change).
  • exhaustive — 5x overfetch, deep graph traversal, no competition suppression. Maximum recall.

get_batch Action (memory tool)

  • memory({ action: "get_batch", ids: ["id1", "id2", ...] }) — retrieve up to 20 full memory nodes in one call.

Changed

  • Token budget raised: 10K → 100K on search and session_context tools.
  • HTTP transport CORS: permissive() → localhost-only origin restriction.
  • Auth token display: Guarded against panic on short tokens.
  • Dormant state threshold: Aligned search (0.3 → 0.4) with memory tool for consistent state classification.
  • cross_reference false positive prevention: Requires 2+ shared words before checking negation signals.

Stats

  • 23 MCP tools, 758 tests passing, 0 failures
  • Full codebase audit: 3 parallel agents, all issues resolved

[2.0.0] - 2026-02-22 — "Cognitive Leap"

The biggest release in Vestige history. A complete visual and cognitive overhaul.

Added

3D Memory Dashboard

  • SvelteKit 2 + Three.js dashboard — full 3D neural visualization at localhost:3927/dashboard
  • 7 interactive pages: Graph (3D force-directed), Memories (browser), Timeline, Feed (real-time events), Explore (connections), Intentions, Stats
  • WebSocket event bus — tokio::broadcast channel with 16 event types (MemoryCreated, SearchPerformed, DreamStarted/Completed, ConsolidationStarted/Completed, RetentionDecayed, ConnectionDiscovered, ActivationSpread, ImportanceScored, Heartbeat, etc.)
  • 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
  • Text label sprites — distance-based visibility (fade in <40 units, out >80 units), canvas-based rendering
  • Dream visualization mode — purple ambient, slow-motion orbit, sequential memory replay
  • FSRS retention curves — SVG R(t) = e^(-t/S) with prediction pills at 1d/7d/30d
  • Command palette — Cmd+K navigation with filtered search
  • Keyboard shortcuts — G Graph, M Memories, T Timeline, F Feed, E Explore, I Intentions, S Stats, / Search
  • Responsive layout — desktop sidebar + mobile bottom nav with safe-area-inset
  • PWA support — installable via manifest.json
  • Single binary deployment — SvelteKit build embedded via include_dir! macro

Engine Upgrades

  • HyDE query expansion — template-based Hypothetical Document Embeddings: classify_intent (6 types) → expand_query (3-5 variants) → centroid_embedding. Wired into semantic_search_raw
  • fastembed 5.11 — upgraded from 5.9, adds Nomic v2 MoE + Qwen3 reranker support
  • Nomic Embed Text v2 MoE — opt-in via --features nomic-v2 (475M params, 305M active, 8 experts, Candle backend)
  • Qwen3 Reranker — opt-in via --features qwen3-reranker (Candle backend, high-precision cross-encoder)
  • Metal GPU acceleration — opt-in via --features metal (Apple Silicon, significantly faster embedding inference)

Backend

  • Axum WebSocket — /ws endpoint with 5-second heartbeat, live stats (memory count, avg retention, uptime)
  • 7 new REST endpoints — POST /api/dream, /api/explore, /api/predict, /api/importance, /api/consolidate, GET /api/search, /api/retention-distribution, /api/intentions
  • Event emission from MCP tools — emit_tool_event() broadcasts events for smart_ingest, search, dream, consolidate, memory, importance_score
  • Shared broadcast channel — single tokio::broadcast::channel(1024) shared between dashboard and MCP server
  • CORS for SvelteKit dev — localhost:5173 allowed in dev mode

Benchmarks

  • Criterion benchmark suite — cosine_similarity 296ns, centroid 1.3µs, HyDE expand 1.4µs, RRF fusion 17µs

Changed

  • Version: 1.8.0 → 2.0.0 (both crates)
  • Rust edition: 2024 (MSRV 1.85)
  • Tests: 651 → 734 (352 core + 378 mcp + 4 doctests)
  • Binary size: ~22MB (includes embedded SvelteKit dashboard)
  • CognitiveEngine moved from main.rs binary crate to lib.rs for dashboard access
  • Dashboard served at /dashboard prefix (legacy HTML kept at / and /graph)
  • McpServer now accepts optional broadcast::Sender<VestigeEvent> for event emission

Technical

  • apps/dashboard/ — new SvelteKit app (Svelte 5, Tailwind CSS 4, Three.js 0.172, @sveltejs/adapter-static)
  • dashboard/events.rs — 16-variant VestigeEvent enum with #[serde(tag = "type", content = "data")]
  • dashboard/websocket.rs — WebSocket upgrade handler with heartbeat + event forwarding
  • dashboard/static_files.rs — include_dir! macro for embedded SvelteKit build
  • search/hyde.rs — HyDE module with intent classification and query expansion
  • benches/search_bench.rs — Criterion benchmarks for search pipeline components

[1.8.0] - 2026-02-21

Added

  • session_context tool — one-call session initialization replacing 5 separate calls (search × 2, intention check, system_status, predict). Token-budgeted responses (~15K tokens → ~500-1000 tokens). Returns assembled markdown context, automationTriggers (needsDream/needsBackup/needsGc), and expandable memory IDs for on-demand retrieval.
  • 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.
  • int8 vector quantization — ScalarKind::F16 → I8 (2x memory savings, <1% recall loss)
  • Migration v7 — FTS5 porter tokenizer (15-30% keyword recall) + page_size 8192 (10-30% faster large-row reads)
  • 22 new tests for session_context and token_budget (335 → 357 mcp tests, 651 total)

Changed

  • Tool count: 18 → 19
  • EmbeddingService::init() changed from &mut self to &self (dead model_loaded field removed)
  • CLAUDE.md updated: session start uses session_context, 19 tools documented, development section reflects storage architecture

Performance

  • Session init: ~15K tokens → ~500-1000 tokens (single tool call)
  • Vector storage: 2x reduction (F16 → I8)
  • Keyword search: 15-30% better recall (FTS5 porter stemming)
  • Large-row reads: 10-30% faster (page_size 8192)
  • Concurrent reads: non-blocking (reader/writer WAL split)

[1.7.0] - 2026-02-20

Changed

  • 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")
  • demote_memory → use memory(action="demote")
  • health_check → use system_status
  • stats → use system_status

[1.6.0] - 2026-02-19

Changed

  • F16 vector quantization — USearch vectors stored as F16 instead of F32 (2x storage savings)
  • Matryoshka 256-dim truncation — embedding dimensions reduced from 768 to 256 (3x embedding storage savings)
  • Convex Combination fusion — replaced RRF with 0.3 keyword / 0.7 semantic weighted fusion for better score preservation
  • Cross-encoder reranker — added Jina Reranker v1 Turbo (fastembed TextRerank) for neural reranking (~20% retrieval quality improvement)
  • Combined: 6x vector storage reduction with better retrieval quality
  • Cross-encoder loads in background — server starts instantly
  • Old 768-dim embeddings auto-migrated on load

[1.5.0] - 2026-02-18

Added

  • CognitiveEngine — 28-module stateful engine with full neuroscience pipeline on every tool call
  • dream tool — memory consolidation via replay, discovers hidden connections and synthesizes insights
  • explore_connections tool — graph traversal with chain, associations, and bridges actions
  • predict tool — proactive retrieval based on context and activity patterns
  • restore tool — restore memories from JSON backup files
  • Automatic consolidation — FSRS-6 decay runs on a 6-hour timer + inline every 100 tool calls
  • ACT-R base-level activation with full access history
  • Episodic-to-semantic auto-merge during consolidation
  • Cross-memory reinforcement on access
  • Park et al. triple retrieval scoring
  • Personalized w20 optimization

Changed

  • All existing tools upgraded with cognitive pre/post processing pipelines
  • Tool count: 19 → 23

[1.3.0] - 2026-02-12

Added

  • importance_score tool — 4-channel neuroscience scoring (novelty, arousal, reward, attention)
  • session_checkpoint tool — batch smart_ingest up to 20 items with Prediction Error Gating
  • find_duplicates tool — cosine similarity clustering with union-find for dedup
  • vestige ingest CLI command for memory ingestion via command line

Changed

  • Tool count: 16 → 19
  • Made get_node_embedding public in core API
  • Added get_all_embeddings for duplicate scanning

[1.2.0] - 2026-02-12

Added

  • Web dashboard — Axum-based on port 3927 with memory browser, search, and system stats
  • memory_timeline tool — browse memories chronologically, grouped by day
  • memory_changelog tool — audit trail of memory state transitions
  • health_check tool — system health status with recommendations
  • consolidate tool — run FSRS-6 maintenance cycle
  • stats tool — full memory system statistics
  • backup tool — create SQLite database backups
  • export tool — export memories as JSON/JSONL with filters
  • gc tool — garbage collect low-retention memories
  • backup_to() and get_recent_state_transitions() storage APIs

Changed

  • Search now supports detail_level (brief/summary/full) to control token usage
  • Tool count: 8 → 16

[1.1.3] - 2026-02-12

Changed

  • Upgraded to Rust edition 2024
  • Security hardening and dependency updates

Fixed

  • Dedup on ingest edge cases
  • Intel Mac CI builds
  • NPM package version alignment
  • Removed dead TypeScript package

[1.1.2] - 2025-01-27

Fixed

  • Embedding model cache now uses platform-appropriate directories instead of polluting project folders
    • macOS: ~/Library/Caches/com.vestige.core/fastembed
    • Linux: ~/.cache/vestige/fastembed
    • Windows: %LOCALAPPDATA%\vestige\cache\fastembed
  • Can still override with FASTEMBED_CACHE_PATH environment variable

[1.1.1] - 2025-01-27

Fixed

  • UTF-8 string slicing issues in keyword search and prospective memory
  • Silent error handling in MCP stdio protocol
  • Feature flag forwarding between crates
  • All GitHub issues resolved (#1, #3, #4)

Added

  • Pre-built binaries for Linux, Windows, and macOS (Intel & ARM)
  • GitHub Actions CI/CD for automated releases

[1.1.0] - 2025-01-26

Changed

  • Tool Consolidation: 29 tools → 8 cognitive primitives
    • recall, semantic_search, hybrid_search → search
    • get_knowledge, delete_knowledge, get_memory_state → memory
    • remember_pattern, remember_decision, get_codebase_context → codebase
    • 5 intention tools → intention
  • Stats and maintenance moved from MCP to CLI (vestige stats, vestige health, etc.)

Added

  • CLI admin commands: vestige stats, vestige health, vestige consolidate, vestige restore
  • Feedback tools: promote_memory, demote_memory
  • 30+ FAQ entries with verified neuroscience claims
  • Storage modes documentation: Global, per-project, multi-Claude household
  • CLAUDE.md templates for proactive memory use
  • Version pinning via git tags

Deprecated

  • Old tool names (still work with warnings, removed in v2.0)

[1.0.0] - 2025-01-25

Added

  • FSRS-6 spaced repetition algorithm with 21 parameters
  • Bjork & Bjork dual-strength memory model (storage + retrieval strength)
  • Local semantic embeddings with fastembed v5 (BGE-base-en-v1.5, 768 dimensions)
  • HNSW vector search with USearch (20x faster than FAISS)
  • Hybrid search combining BM25 keyword + semantic + RRF fusion
  • Two-stage retrieval with reranking (+15-20% precision)
  • MCP server for Claude Desktop integration
  • Tauri desktop application
  • Codebase memory module for AI code understanding
  • Neuroscience-inspired memory mechanisms:
    • Synaptic Tagging and Capture (retroactive importance)
    • Context-Dependent Memory (Tulving encoding specificity)
    • Spreading Activation Networks
    • Memory States (Active/Dormant/Silent/Unavailable)
    • Multi-channel Importance Signals (Novelty/Arousal/Reward/Attention)
    • Hippocampal Indexing (Teyler & Rudy 2007)
  • Prospective memory (intentions and reminders)
  • Sleep consolidation with 5-stage processing
  • Memory compression for long-term storage
  • Cross-project learning for universal patterns

Changed

  • Upgraded embedding model from all-MiniLM-L6-v2 (384d) to BGE-base-en-v1.5 (768d)
  • Upgraded fastembed from v4 to v5

Fixed

  • SQL injection protection in FTS5 queries
  • Infinite loop prevention in file watcher
  • SIGSEGV crash in vector index (reserve before add)
  • Memory safety with Mutex wrapper for embedding model

[0.1.0] - 2025-01-24

Added

  • Initial release
  • Core memory storage with SQLite + FTS5
  • Basic FSRS scheduling
  • MCP protocol support
  • Desktop app skeleton