@andespindola/brainlink 0.1.0-beta.11 → 0.1.0-beta.110

package/AGENTS.md

@@ -6,19 +6,19 @@ This file tells coding agents and AI assistants how to use this repository.
 
 Brainlink is a local-first knowledge memory for agents.
 
-It reads a Markdown vault, extracts `[[wiki links]]` and `#tags`, builds a local SQLite full-text index, and returns compact context packages that agents can inject into prompts.
+It reads a Markdown vault, extracts `[[wiki links]]` and `#tags`, builds a local file index at `.brainlink/index.json`, and returns compact context packages that agents can inject into prompts.
 
 ## Source Of Truth
 
 Markdown files are the source of truth.
 
-The SQLite database at `.brainlink/brainlink.db` is a derived index. It can be deleted and rebuilt with:
+The JSON index at `.brainlink/index.json` is derived. It can be deleted and rebuilt with:
 
 ```bash
 npm run dev -- index --vault ./vault
 ```
 
-Do not store permanent knowledge only in SQLite.
+Do not store permanent knowledge only in index artifacts.
 
 By default, the installed Brainlink CLI uses `$HOME/.brainlink/vault` as its vault. Passing `--vault` or setting `vault` in `brainlink.config.json` intentionally selects a custom vault such as `./vault`.
 
@@ -83,6 +83,9 @@ Use watch mode while editing notes:
 ```bash
 npm run dev -- server --vault ./vault --watch
 npm run dev -- watch --vault ./vault
+npm run dev -- bench --vault ./vault
+npm run dev -- bench --vault ./vault --watch
+npm run dev -- pack-backup --vault ./vault
 ```
 
 Start MCP over stdio:
@@ -107,10 +110,10 @@ npm run dev -- doctor --vault ./vault
 
 - Keep domain rules in `src/domain`.
 - Keep use cases in `src/application`.
-- Keep filesystem and SQLite details in `src/infrastructure`.
+- Keep filesystem and index details in `src/infrastructure`.
 - Keep CLI concerns in `src/cli`.
 - Prefer pure functions for parsing, ranking, formatting, and transformation.
-- Do not make SQLite the canonical storage layer.
+- Do not make index artifacts the canonical storage layer.
 - Do not add comments with emojis.
 - Keep JSON output backwards compatible where possible.
 

package/CHANGELOG.md

@@ -22,6 +22,30 @@
 - Added short-lived hybrid search cache with automatic invalidation on index changes.
 - Added `stats --extended` observability output with storage, quality and latency probes.
 - Added `docs/QUICKSTART.md` and aligned README/agent docs with the latest CLI/MCP flows.
+- Added middle-out context assembly so chunk selection expands around the strongest note chunk.
+- Added compressed-space pack prefiltering (token bloom index) before `.blpk` decryption and scan.
+- Improved graph UI auto-fit and viewport recovery so loaded nodes are re-centered when zoom/pan drifts to empty canvas.
+- Added cross-platform native desktop GUI auto-open for `blink server` (macOS Swift/WebKit, Windows PowerShell WinForms, Linux Python GTK/WebKit2), with app-window/browser fallback.
+- Changed Linux default UI launch to app-window/browser for lighter startup; Linux native GUI is now opt-in via `BRAINLINK_LINUX_NATIVE_GUI=1`.
+- Added native GUI parent-process monitoring so GUI windows close automatically when `blink server` stops.
+- Improved non-mac browser detection fallback to try installed Edge/Chrome/Firefox/Chromium candidates before system default open.
+- Improved graph filter rendering to keep hub anchor nodes visible (`Memory Hub`/`MOC`/high-degree fallback) for coherent relationship context.
+- Fixed graph modal content loading by correcting agent query parameter composition for `/api/graph-node` and `/api/graph-filter` requests.
+- Improved 50k+ graph rendering performance with viewport-aware spatial node culling, cached render visibility, and node-adjacent edge selection to avoid full graph scans every frame.
+- Added incremental vault indexing with file snapshots to reuse unchanged documents/chunks/embeddings, plus adaptive search-pack rebuild thresholds to avoid full re-compression on small edits.
+- Reduced large-graph HTTP payload size with compact `/api/graph-layout` encoding for high-node vaults and capped transmitted edges to improve UI load responsiveness.
+- Added aggressive graph LOD clustering when zoomed out, dynamic per-zoom edge render budgets, and a dedicated frontend worker for off-main-thread graph filter matching.
+- Improved Linux browser fallback launch stability by auto-applying Chromium compatibility flags (`--ozone-platform=x11`, `--disable-gpu`, `--disable-features=Vulkan,VaapiVideoDecoder`, `--disable-background-networking`) for app-window/browser modes.
+- Improved massive-graph UI responsiveness with stricter render budgets, adaptive heavy-graph frame throttling, reduced interaction hit-test frequency, and URL-first agent selection on initial graph load.
+- Improved 50k+ graph LOD behavior so zoomed-out views render lightweight cluster overviews and progressively reveal nodes/edges only as zoom increases.
+- Added `blink bench` with realtime index phase telemetry and per-run compressed-pack analysis (input/output bytes, ratio, saved space, rebuild reason and duration), including continuous watch mode.
+- Added tunable single-stage search-pack compression settings (`searchPack.rowChunkSize`, `searchPack.compressionLevel`, `searchPack.useDictionary`).
+- Added benchmark guardrails for compression savings and latency regression (`searchPack.guardrailMinSavingsPercent`, `searchPack.guardrailMaxLatencyRegressionPercent`), reported in `blink bench`.
+- Added `blink pack-backup` for offline second-stage compression backups of encrypted `.blpk` packs, outside the online query path.
+- Hardened Linux browser launch flags for Ubuntu 26 Chromium/Wayland compatibility (`--disable-vulkan`, `--use-gl=swiftshader`, `--ozone-platform-hint=x11`).
+- Improved pack resilience by auto-repairing missing search-pack manifests from existing `.blpk` files, avoiding unnecessary full repacks on small incremental updates.
+- Updated Linux graph auto-open behavior to prioritize the system default browser (`xdg-open`) before explicit browser fallbacks.
+- Removed implicit Chromium dependency in Linux auto-open flow; app-window launch is now opt-in (`BRAINLINK_LINUX_APP_WINDOW=1`).
 
 ## 0.1.0-beta.3
 
@@ -49,8 +73,8 @@
 ## 0.1.0-alpha.0
 
 - Added local-first Markdown vault indexing.
-- Added SQLite FTS, local semantic retrieval, wiki links, backlinks and graph retrieval.
-- Added SQLite semantic bucket indexing to narrow vector candidates for larger vaults.
+- Added local full-text indexing, local semantic retrieval, wiki links, backlinks and graph retrieval.
+- Added semantic candidate bucket indexing to narrow vector candidates for larger vaults.
 - Optimized title/link resolution with precomputed agent-scoped title maps.
 - Added CLI, JSON output, HTTP API and graph UI.
 - Added vault diagnostics: stats, broken links, orphans, validation and doctor.

package/CONTRIBUTING.md

@@ -22,7 +22,7 @@ npm run pack:smoke
 ## Design Rules
 
 - Markdown files are the source of truth.
-- SQLite is a derived index and must remain rebuildable.
+- Local index artifacts are derived and must remain rebuildable.
 - Domain parsing, graph analysis and layout should stay pure and testable.
-- CLI, HTTP, filesystem and SQLite code are adapters around application use cases.
+- CLI, HTTP, filesystem and index code are adapters around application use cases.
 - MCP integration should live outside this package by wrapping the CLI with `--json`.

package/COPYRIGHT.md

@@ -0,0 +1,5 @@
+Copyright (c) 2026 Substructa
+
+This project is licensed under the MIT License.
+
+See [LICENSE](./LICENSE) for full terms.

package/README.md

@@ -52,13 +52,15 @@ LLMs do not have infinite context. Brainlink gives agents an external memory lay
 1. Durable knowledge is written as Markdown.
 2. Notes are connected with `[[wiki links]]`.
 3. Concepts are classified with `#tags`.
-4. Brainlink builds a local SQLite index with FTS records and local embeddings.
+4. Brainlink builds a local JSON index (`.brainlink/index.json`) and private encrypted search packs.
 5. Agents query the index before responding.
 6. Brainlink returns compact, source-backed context.
 
-Markdown is the source of truth. `.brainlink/brainlink.db` is only a rebuildable index.
-Brainlink now keeps an automatic rollback snapshot at `.brainlink/brainlink.db.backup`. If the main SQLite file is corrupted, Brainlink automatically restores from snapshot (or recreates a clean index when no snapshot exists).
-After each index run, Brainlink also writes compressed search packs at `.brainlink/search-packs/*.jsonl.gz`. If SQLite is unavailable, search falls back to these packs automatically.
+Markdown is the source of truth. `.brainlink/index.json` is a rebuildable index artifact.
+After each index run, Brainlink also writes private encrypted search packs at `.brainlink/search-packs/*.blpk` to preserve fast retrieval and portable recovery.
+Online retrieval always uses a single compression stage per pack; optional second-stage compression is reserved for offline backup artifacts only.
+Pack decryption uses a Brainlink key from `$BRAINLINK_HOME/keys` or from `BRAINLINK_SEARCH_PACK_KEY` when explicitly configured.
+Legacy `.jsonl.gz` packs are upgraded to `.blpk` automatically on first search/context access.
 
 ## Features
 
@@ -66,8 +68,12 @@ After each index run, Brainlink also writes compressed search packs at `.brainli
 - Obsidian-compatible `[[wiki links]]` and `#tags`.
 - Weighted graph edges so agents can rank relationship importance and priority.
 - Backlinks, broken-link reports, orphan detection and validation.
-- Full-text, semantic and hybrid retrieval modes.
-- SQLite-backed semantic candidate buckets for larger vaults.
+- Full-text, semantic and hybrid retrieval on a local file index.
+- Middle-out context assembly around the strongest chunk per document.
+- In-process index and context caching with automatic invalidation on index updates.
+- Compressed-space prefiltering for `.blpk` packs before decryption and scan.
+- Incremental indexing that reprocesses only changed markdown files and reuses existing chunks/embeddings for unchanged notes.
+- Adaptive compressed-pack rebuild policy to keep indexing fast during small edit batches.
 - Agent namespaces under `agents/<agent-id>/`.
 - S3-compatible bucket vaults through `s3://bucket/prefix` URIs.
 - CLI with machine-readable `--json` output.
@@ -75,6 +81,17 @@ After each index run, Brainlink also writes compressed search packs at `.brainli
 - Built-in MCP stdio server for agent tool integration.
 - Local HTTP API.
 - Realtime graph UI with agent selector and colored knowledge groups.
+- Graph renderer optimized for large datasets with viewport-driven node culling and edge lookup by visible nodes.
+- Canvas graph rendering uses the same batched node and edge pipeline for every graph size, reducing per-frame draw calls while keeping selected and hovered items highlighted.
+- WebGL acceleration is used when available for dense node and edge drawing, with Canvas 2D preserved as the interaction and fallback layer.
+- Graph zoom-out renders hierarchical ecosystem subgraphs only above 1000 notes: the memory hub stays centered, 1000-note groups stay as compact sand-like points, and focused groups gradually expand into smaller graph meshes before individual notes are rendered.
+- Large graph layout API automatically uses compact payload encoding with link-coverage-aware edge selection to reduce initial client load without hiding major relationships.
+- Large-segment layout spacing now grows logarithmically to keep initial visual density consistent between medium and very large vaults (for example, ~1k vs ~50k notes).
+- Graph coordinates are visually compacted across graph sizes so reset starts from a stable macro mass and zoom-in progressively expands toward local detail.
+- Zoomed-out graph LOD renders nested subgraphs and progressively expands only the focused cluster as zoom increases, including very large vaults.
+- Graph reset starts in macro "galaxy" overview mode and progressively reveals nearby nodes as zoom increases, including smaller vaults.
+- Graph filtering runs in a dedicated browser worker to keep the UI thread responsive during heavy datasets.
+- Edge rendering budgets adapt to zoom level to prevent frame spikes on large graph panoramas.
 
 ## Install
 
@@ -284,7 +301,7 @@ export BRAINLINK_S3_FORCE_PATH_STYLE=1
 
 Bucket vaults mirror Markdown into a local cache under
 `$BRAINLINK_HOME/bucket-cache`. The bucket remains canonical; the local
-`.brainlink/brainlink.db` stays a disposable index. Run `index` after remote
+`.brainlink/index.json` stays a disposable index artifact. Run `index` after remote
 bucket changes before relying on `search`, `context`, graph or validation
 commands. Watch mode is only supported for local filesystem vaults.
 
@@ -300,7 +317,7 @@ vault/
     research-agent/
       source-review-policy.md
   .brainlink/
-    brainlink.db
+    index.json
 ```
 
 Permanent data:
@@ -310,7 +327,7 @@ Permanent data:
 
 Rebuildable data:
 
-- `.brainlink/brainlink.db`
+- `.brainlink/index.json`
 - full-text records
 - local embedding vectors
 - local embedding buckets
@@ -394,6 +411,7 @@ blink agent upgrade
 ```
 
 This configures `~/.codex/config.toml` with Brainlink MCP (`brainlink-mcp`) so Brainlink is available by default in agent sessions.
+`agent install` and `agent upgrade` also apply the MCP `fully-auto` bootstrap policy by default (`enforceBootstrap`, `enforceContextFirst`, `autoBootstrapOnRead`, `autoBootstrapOnStartup` all enabled).
 
 If you are inside this repository and want plugin gallery setup too:
 
@@ -513,6 +531,8 @@ Available tools:
 - `brainlink_recommendations`: return an automatic action plan so agents can run Brainlink in the recommended order.
 - `brainlink_context`: read indexed context for a task or question.
 - `brainlink_search`: search indexed notes.
+- `brainlink_dedupe`: detect duplicate candidates using exact hash + semantic similarity scores.
+- `brainlink_resolve_duplicate`: resolve duplicate pairs (`merge`, `link`, `ignore`) with connectivity-safe fallback edges.
 - `brainlink_add_note`: write durable Markdown memory and reindex.
 - `brainlink_add_file`: ingest a local file as a note and reindex.
 - `brainlink_index`: rebuild the vault index.
@@ -551,6 +571,17 @@ blink server --host 127.0.0.1 --port 4321 --watch
 ```
 
 By default, the server uses `$HOME/.brainlink/vault`. Pass `--vault ./vault` only when you want to inspect a custom vault.
+By default, `blink server` tries to open the graph in a native desktop GUI window:
+- macOS: Swift + WebKit
+- Windows: PowerShell WinForms WebBrowser
+- Linux: optional Python GTK + WebKit2 (requires `python3` + `gi` + `WebKit2`)
+
+On Linux, native GUI is disabled by default for better startup performance. Enable it with `BRAINLINK_LINUX_NATIVE_GUI=1`.
+If native GUI launch is unavailable on your system, it falls back to dedicated app-window mode and then to the default browser.
+For Chromium-family browsers on Linux (`chromium`, `chromium-browser`, `google-chrome`, `microsoft-edge`, `brave-browser`), Brainlink now auto-applies compatibility flags during launch (`--ozone-platform=x11`, `--ozone-platform-hint=x11`, `--disable-gpu`, `--disable-vulkan`, `--use-gl=swiftshader`, `--disable-features=Vulkan,VaapiVideoDecoder`, `--disable-background-networking`) to avoid common Wayland/Vulkan/VAAPI startup issues.
+On Linux, Brainlink opens the graph through the system default browser first (`xdg-open`), then `$BROWSER`/detected browsers as fallback. Chromium-family app-window mode is optional via `BRAINLINK_LINUX_APP_WINDOW=1`.
+Use `--no-open` to keep it headless.
+When native GUI is used, the GUI window automatically closes when the `blink server` process stops.
 
 The graph UI shows:
 
@@ -558,13 +589,26 @@ The graph UI shows:
 - `[[wiki links]]` as weighted edges
 - details opened on node click (tags, outgoing links, backlinks, full Markdown content)
 - neutral graph nodes with segment/group metadata
-- agent selector for isolated views
+- agent selector (id-only labels) for isolated views
 - graph filter matches title, path, tags and note content
+- graph filter keeps hub context nodes visible (`Memory Hub`/`MOC`/high-degree fallback) to preserve relationship readability
 - realtime refresh while `--watch` is enabled
 - graph controls for zoom in, zoom out, fit visible nodes and reset-to-fit-all
-- wheel zoom anchored to cursor position for faster navigation in large graphs
+- wheel zoom (including `cmd+scroll` and `ctrl+scroll`) anchored to cursor position for faster navigation in large graphs
+- zoom-out floor for large and massive graphs, plus reset macro floor tied to hub-neighbor distance and first-level cluster spacing, with a tighter initial camera so the first graph appears as a closer cell instead of a distant mesh
+- keyboard shortcuts: `+` zoom in, `-` zoom out, `0` reset fit
+- double-click on canvas zooms in at cursor position
 - floating graph totals (notes, links, tags) below the Brainlink title
-- large-graph rendering safeguards (edge draw caps, lower redraw rate, zoom-aware interaction)
+- graph rendering safeguards (batched canvas drawing across graph sizes, edge draw caps, lower redraw rate, zoom-aware interaction)
+- adaptive CPU safeguards for large graphs: idle frame pacing, throttled background physics updates and cached viewport dimensions to reduce redraw/layout overhead while preserving interaction responsiveness
+- hierarchical hot-path optimizations reduce per-frame allocations and repeated scans during layered cluster expansion and edge projection
+- hierarchical edge projection now caches hub membership and node-to-cluster resolution per frame to keep large recursive subgraph rendering smooth during continuous zoom and pan
+- hierarchical projection now uses stronger perspective yaw/pitch and depth-based render ordering so layered subgraphs read as a true 3D field instead of a flat expansion
+- node rendering also applies depth projection cues on large visible sets (position, scale, opacity and edge depth weighting) so the graph keeps 3D perception when leaving cluster-only LOD layers
+- WebGL node and edge acceleration when supported, falling back to Canvas 2D without changing graph behavior
+- compact macro-to-micro density progression so reset keeps the graph mass oriented and zoom-in separates local neighborhoods progressively
+- graph camera treats hub-centered navigation as structural only when the hub is dominant; diffuse stress graphs reset and zoom around the full graph mass
+- graph LOD progression: hierarchical rendering now follows one recursive graph-of-graphs standard whenever a graph has more than one hierarchy level; each level expands through intermediate subgraph sizes (instead of jumping directly to leaves), starts from a memory-hub-centered mesh, and each supernode can expand into another same-shape subgraph level (up to 999 children) with latent fade-in, aggregated real links and local sibling mesh links so org-heavy-like and stress-50k-like structures share the same layered behavior at different depths; layered clusters also receive stronger perspective depth projection (Z-depth) with vertical camera tilt/parallax so expansion reads as a true depth field instead of a flat 2D switch; for massive graphs the first expansion starts deeper in zoom and is additionally gated by focus readiness (screen-space isolation of the focused parent) so child levels open only when that subgraph is truly centered and separated in view
 
 The server indexes before starting by default. Use `--no-index` to skip that step:
 
@@ -666,6 +710,18 @@ blink migrate-vault --from ~/.brainlink/vault --to ./team-vault --report ./migra
 Runs explicit markdown migration between vaults while preserving conflicts as `.conflict-<timestamp>` files.  
 Use `--dry-run` to preview `copied`, `conflicted` and `unchanged` counts before writing.
 
+### `db-import`
+
+```bash
+blink db-import --vault ./team-vault
+blink db-import --vault ./team-vault --db ./legacy/brainlink.db
+blink db-import --vault ./team-vault --db ./legacy/brainlink.db --table legacy_notes --dry-run
+```
+
+Imports durable memory from a legacy SQLite database into Markdown notes (`agents/<agent-id>/*.md`) and reindexes by default.
+When `--db` is omitted, Brainlink auto-detects common legacy paths such as `<vault>/.brainlink/brainlink.db`.
+Use `--agent <id>` to force all imported rows into one namespace, `--limit` for incremental imports, `--dry-run` to preview without writing files, and `--no-index` to defer reindexing.
+
 ### `init`
 
 ```bash
@@ -690,6 +746,28 @@ blink add "Note Title" --vault ./vault --content-file ./notes.md --no-auto-index
 
 Creates a Markdown note under `agents/<agent-id>/`. Common secret patterns are blocked by default; use `--allow-sensitive` only for an intentionally protected vault.
 To avoid disconnected memory, Brainlink auto-adds a fallback wiki edge when a note is written without links, creating agent hub notes when needed.
+`add` also returns `possibleDuplicates` (exact hash + semantic candidates) so agents can resolve duplicate memory right after writes.
+
+### `dedupe`
+
+```bash
+blink dedupe --vault ./vault --json
+blink dedupe --vault ./vault --agent coding-agent --limit 20 --min-score 0.92 --json
+blink dedupe --vault ./vault --no-semantic --json
+```
+
+Detects `possibleDuplicate` pairs using exact content hashes and optional semantic similarity.
+
+### `dedupe-resolve`
+
+```bash
+blink dedupe-resolve --vault ./vault --left agents/shared/a.md --right agents/shared/b.md --action merge --json
+blink dedupe-resolve --vault ./vault --left agents/shared/a.md --right agents/shared/b.md --action link --json
+blink dedupe-resolve --vault ./vault --left agents/shared/a.md --right agents/shared/b.md --action ignore --json
+```
+
+Resolves a duplicate pair with `merge`, `link` or `ignore`.
+When action is not `merge`, Brainlink still creates a low-priority related edge (`#related-to`) so notes remain connected.
 
 ### `index`
 
@@ -700,6 +778,38 @@ blink index --vault ./vault
 
 Rebuilds the local index from Markdown files.
 
+### `bench`
+
+```bash
+blink bench --vault ./vault
+blink bench --vault ./vault --watch
+blink bench --vault ./vault --watch --debounce 500
+blink bench --vault ./vault --json
+```
+
+Runs indexing with realtime phase telemetry (`start`, `scan`, `parse`, `embed`, `persist`, `packs`, `complete`) and prints a benchmark summary at the end of each run.
+
+Summary includes compression behavior for `.blpk` packs when rebuild happens:
+- pack rebuild reason
+- pack count and pack build duration
+- uncompressed input bytes vs compressed output bytes
+- saved percentage
+- objective guardrails (minimum savings and maximum latency regression thresholds)
+
+Use `--watch` to keep benchmarking incremental reindex runs after Markdown changes (local filesystem vaults only).
+When `.brainlink/search-packs/manifest.json` is missing but `.blpk` files exist, Brainlink repairs the manifest first and avoids unnecessary full pack rebuild on small edits.
+
+### `pack-backup`
+
+```bash
+blink pack-backup --vault ./vault
+blink pack-backup --vault ./vault --output ./vault/.brainlink/backups/custom.blpkbak.gz
+blink pack-backup --vault ./vault --json
+```
+
+Creates an offline backup artifact of encrypted search packs with a second compression pass.
+This is intentionally outside the online retrieval path (`index`, `search`, `context`).
+
 ### `agents`
 
 ```bash
@@ -722,11 +832,12 @@ If `--mode` or `--limit` is omitted, Brainlink resolves values from the current
 
 Modes:
 
-- `hybrid`: default; combines SQLite FTS with local embedding similarity.
-- `fts`: exact lexical retrieval through SQLite FTS.
+- `hybrid`: default; combines lexical matching with local embedding similarity.
+- `fts`: exact lexical retrieval from the file index.
 - `semantic`: local deterministic embedding similarity only.
 
 Hybrid results are cached in-memory for a short TTL and invalidated automatically when the local index file changes.
+Context selection uses a middle-out strategy: it starts from the strongest chunk in a note and expands to neighboring chunks while respecting token budget.
 
 ### `context`
 
@@ -737,6 +848,7 @@ blink context "question" --vault ./vault --agent coding-agent --mode hybrid --js
 ```
 
 Builds a compact context package for an agent.
+Repeated calls with the same vault, agent, query, mode and token/limit settings are served from a short in-memory cache while the index is unchanged.
 
 ### `links`
 
@@ -821,9 +933,15 @@ Watches Markdown files and rebuilds the index when notes change.
 ```bash
 blink server --watch
 blink server --vault ./vault --watch
+blink server --vault ./vault --watch --no-open
 ```
 
 Starts the local read-only graph UI and HTTP API.
+By default, it tries to open a native desktop GUI window for the graph URL.
+On Linux, native GUI is disabled by default; enable it with `BRAINLINK_LINUX_NATIVE_GUI=1`.
+If native GUI launch is unavailable, it falls back to dedicated app-window mode and then browser open.
+When fallback opens Chromium-family browsers on Linux, Brainlink automatically uses compatibility launch flags for stable rendering on Ubuntu/Wayland setups.
+Use `--no-open` to skip that behavior.
 
 The HTTP server only binds to loopback hosts such as `127.0.0.1`, `localhost` or `::1`.
 
@@ -864,6 +982,13 @@ If no `vault` is configured and no `--vault` flag is passed, Brainlink uses `$HO
   "embeddingProvider": "local",
   "defaultSearchMode": "hybrid",
   "chunkSize": 1200,
+  "searchPack": {
+    "rowChunkSize": 5000,
+    "compressionLevel": 5,
+    "useDictionary": true,
+    "guardrailMinSavingsPercent": 8,
+    "guardrailMaxLatencyRegressionPercent": 5
+  },
   "agentProfiles": {
     "coding-agent": {
       "defaultSearchMode": "semantic",
@@ -975,7 +1100,7 @@ src/
   application/      use cases
   cli/              command-line adapter
   domain/           pure knowledge rules
-  infrastructure/   filesystem and SQLite adapters
+  infrastructure/   filesystem and index adapters
 ```
 
 Detailed notes:
@@ -987,7 +1112,6 @@ Detailed notes:
 ## Current Limits
 
 - Semantic search uses deterministic local embeddings, not a remote model provider.
-- Semantic search uses SQLite embedding buckets to narrow candidates before cosine scoring.
 - `embeddingProvider` currently supports `local` and `none`.
 - Link resolution is title-based inside each agent namespace, with `shared` as fallback.
 - HTTP API is local and unauthenticated.
@@ -998,7 +1122,7 @@ Detailed notes:
 The `0.1.0-beta` line is intended to stabilize the local-first memory loop:
 
 - Markdown as durable memory.
-- SQLite FTS plus local embeddings and semantic buckets as rebuildable retrieval index.
+- Rebuildable file index plus local embeddings and encrypted pack exports.
 - CLI as the primary agent interface.
 - HTTP graph API and frontend as inspection tools.
 - Agent namespaces to avoid context mixing.
@@ -1014,7 +1138,7 @@ Brainlink is local-first by default.
 - Brainlink HTTP is localhost-only and refuses non-loopback hosts.
 - Brainlink blocks common secret patterns by default when adding notes. Use `--allow-sensitive` only for intentional, protected vaults.
 - Do not store secrets, credentials, API keys or regulated personal data unless the vault is protected by your own storage controls.
-- Treat `.brainlink/brainlink.db` as disposable derived data.
+- Treat `.brainlink/index.json` and `.brainlink/search-packs/` as disposable derived artifacts.
 
 See [SECURITY.md](SECURITY.md).
 
@@ -1025,6 +1149,7 @@ See [CONTRIBUTING.md](CONTRIBUTING.md).
 ## License
 
 MIT. See [LICENSE](LICENSE).
+Copyright (c) 2026 Substructa. See [COPYRIGHT.md](COPYRIGHT.md).
 
 ### Memory Optimization Loop (1-7)
 

package/SECURITY.md

@@ -7,7 +7,7 @@ Brainlink is local-first.
 - The HTTP server binds to `127.0.0.1` by default.
 - The HTTP server always refuses non-loopback hosts.
 - The HTTP server is read-only and does not expose note creation, indexing or update routes.
-- The SQLite database is a derived local index.
+- Local index artifacts (`.brainlink/index.json` and `.brainlink/search-packs/`) are derived data.
 - Markdown files are user-owned source data.
 - Brainlink-created Markdown files use `0600` permissions.
 - Brainlink-created directories and `.brainlink` use `0700` permissions.

package/dist/application/analyze-vault.js

@@ -1,7 +1,5 @@
 import { stat } from 'node:fs/promises';
-import { existsSync } from 'node:fs';
 import { performance } from 'node:perf_hooks';
-import { join } from 'node:path';
 import { validateGraph, getBrokenLinks, getOrphanNodes, getVaultStats } from '../domain/graph-analysis.js';
 import { ensureVault, listVaultFiles, readMarkdownFiles } from '../infrastructure/file-system-vault.js';
 import { resolveAgentRuntimeDefaults } from '../infrastructure/config.js';
@@ -96,17 +94,11 @@ export const doctorVault = async (vaultPath) => {
     const files = await readMarkdownFiles(absoluteVaultPath);
     const graph = await getGraphSummary(absoluteVaultPath);
     const validation = validateGraph(graph);
-    const backupPath = join(absoluteVaultPath, '.brainlink', 'brainlink.db.backup');
-    const hasBackup = existsSync(backupPath);
-    const backupReady = graph.nodes.length === 0 || hasBackup;
     const checks = [
         createCheck('vault', true, `Vault ready at ${absoluteVaultPath}`),
         createCheck('markdown-files', files.length > 0, `${files.length} markdown files found`),
         createCheck('index', graph.nodes.length > 0, `${graph.nodes.length} indexed documents found`),
-        createCheck('broken-links', validation.brokenLinks.length === 0, `${validation.brokenLinks.length} broken links found`),
-        createCheck('index-backup', backupReady, backupReady
-            ? (hasBackup ? 'SQLite recovery snapshot is available' : 'No index yet. Snapshot will be created after first indexing run')
-            : 'Recovery snapshot missing. Run blink index to create a rollback snapshot')
+        createCheck('broken-links', validation.brokenLinks.length === 0, `${validation.brokenLinks.length} broken links found`)
     ];
     const recommendations = files.length === 0 && graph.nodes.length === 0
         ? [

package/dist/application/build-context.js

@@ -1,13 +1,68 @@
+import { stat } from 'node:fs/promises';
 import { formatContextPackage, selectContextSections } from '../domain/context.js';
+import { indexStoragePath } from '../infrastructure/file-index.js';
 import { searchKnowledge } from './search-knowledge.js';
+const contextCacheTtlMs = 45_000;
+const contextCacheMaxEntries = 200;
+const contextCache = new Map();
+const readIndexMtimeMs = async (vaultPath) => {
+    try {
+        return (await stat(indexStoragePath(vaultPath))).mtimeMs;
+    }
+    catch {
+        return 0;
+    }
+};
+const toCacheKey = (vaultPath, query, limit, maxTokens, agentId, mode) => JSON.stringify({
+    vaultPath,
+    query: query.trim().toLowerCase(),
+    limit,
+    maxTokens,
+    agentId: agentId?.trim().toLowerCase() ?? '*',
+    mode: mode ?? 'default'
+});
+const contextCacheGet = (key, indexMtimeMs) => {
+    const entry = contextCache.get(key);
+    if (!entry) {
+        return undefined;
+    }
+    const fresh = Date.now() - entry.createdAt <= contextCacheTtlMs && entry.indexMtimeMs === indexMtimeMs;
+    if (!fresh) {
+        contextCache.delete(key);
+        return undefined;
+    }
+    return entry.context;
+};
+const contextCacheSet = (entry) => {
+    contextCache.set(entry.key, entry);
+    if (contextCache.size <= contextCacheMaxEntries) {
+        return;
+    }
+    const overflow = contextCache.size - contextCacheMaxEntries;
+    const keys = Array.from(contextCache.keys()).slice(0, overflow);
+    keys.forEach((key) => contextCache.delete(key));
+};
 export const buildContextPackage = async (vaultPath, query, limit, maxTokens, agentId, mode) => {
+    const cacheKey = toCacheKey(vaultPath, query, limit, maxTokens, agentId, mode);
+    const indexMtimeMs = await readIndexMtimeMs(vaultPath);
+    const cached = contextCacheGet(cacheKey, indexMtimeMs);
+    if (cached) {
+        return cached;
+    }
     const results = await searchKnowledge(vaultPath, query, limit, agentId, mode);
     const sections = selectContextSections(results, maxTokens);
-    return {
+    const context = {
         query,
         sections,
         content: formatContextPackage(query, sections)
     };
+    contextCacheSet({
+        key: cacheKey,
+        createdAt: Date.now(),
+        indexMtimeMs,
+        context
+    });
+    return context;
 };
 export const buildContext = async (vaultPath, query, limit, maxTokens, agentId, mode) => {
     const contextPackage = await buildContextPackage(vaultPath, query, limit, maxTokens, agentId, mode);