Exocortex v0.5.0

What's New

Knowledge Wiki System (Karpathy-inspired)

• Wiki compilation — compile memories into interlinked markdown articles
• Incremental wiki refresh — only recompile stale articles (memory_wiki_refresh)
• Obsidian live sync — incrementally sync memories to an Obsidian vault (memory_obsidian_sync)
• Query-to-wiki promotion — persist analyses as wiki articles (memory_promote)
• Knowledge lint — one-command health audit: contradictions, stale claims, orphans, suggested topics (memory_lint)
• Web clipper — one-click URL ingest via MCP or REST (memory_clip, POST /api/library/clip)

Retrieval Improvements

• 97.0% Recall@5 on LongMemEval (up from prior baseline)
• Two-pass FTS search — AND then OR strategy for better precision
• Auto query expansion — server-side domain synonym expansion for embeddings
• LLM re-ranking — optional re-rank pass on top-N results (opt-in via search.rerank_enabled)
• Retrieval regression gate — 7 golden queries including production friction queries

New MCP Tools

• memory_lint — comprehensive knowledge-base audit
• memory_wiki_refresh — incremental wiki update
• memory_promote — promote analysis to wiki article
• memory_clip — web clipper URL ingest
• memory_correct — correct a memory with supersession chain
• memory_job_health — job outcome tracking with alert thresholds
• memory_obsidian_sync — incremental Obsidian vault sync

Import & Export

• Obsidian vault import — parses YAML frontmatter, wikilinks, infers tiers from folder structure
• ChatGPT/Claude import — auto-tags by topic, weights longer conversations higher
• Source immutability — ingested URLs flagged immutable, excluded from consolidation

Intelligence

• Temporal expiry detection — auto-expires memories with "tomorrow", "by Friday", "in 3 days"
• Auto entity categorization — technologies (≥0.9) and organizations (≥0.85) auto-typed
• LLM consolidation summaries — optional LLMSummarizer interface for higher-quality merges
• Provenance chain tracking — derivation depth, source trust, origin pipeline
• Content hash dedup — prevents duplicate consolidation summaries

Security

• Path traversal fixes, SQL injection hardening
• x402 replay attack prevention
• SAVEPOINT-based transaction nesting

Performance

• MCP tools consolidated 38 → 21
• Prepared statement caching, batch embeddings
• Embedding model selector (ConvoMem benchmark)

v0.4.0

What's New

Retrieval Intelligence

• Scoring weight optimizer — coordinate descent auto-tunes vector, FTS, recency, graph, and usefulness weights
• Quality-tiered storage — persisted quality scores, auto-triage, embedding upgrades
• Goal-gated retrieval — active goals boost relevant memories in search

Memory Lifecycle

• Consolidation quality gate — agent-driven summaries replace mechanical merges
• Self-audit alerts — surfaced at session start for proactive maintenance
• Prediction & calibration system — track forecasts and measure accuracy over time

Obsidian Export

• Curated topic documents with cross-links and project enrichment
• Technical principles and project map sections
• Auto-detect projects from memory graph

Dashboard

• Knowledge graph improvements (community detection, centrality analysis)
• Empty states, stats cards, recent filter, library search
• Community summaries and memory evolution views

Infrastructure

• Security hardening — auth, CSP, rate limiting, path traversal protection
• ChatGPT & Claude conversation import
• Research rate limiting — paces DuckDuckGo and URL ingestion
• N+1 query fixes, new indexes, graph endpoint optimization

43 commits since v0.3.0.

v0.3.0

Changes

• Update screenshots with richer test data and milestones
• Update README with new tools, routes, and analytics; fix importance test
• feat: add dedup exempt tags, golden queries, and contradiction bulk dismiss
• feat: dashboard round 5 — knowledge gaps, regression UI, reembed/consolidate buttons, mini link graph
• feat: improve compact previews, add contradictions tool, filter session-orient by relevance
• feat: round 3-4 improvements — scoring, analytics, dashboard UX
• fix: graph drag, search freeze, and hover hit detection
• feat: interactive knowledge graph, decision timeline fix, search improvements
• feat: query expansion, analytics dashboard, backup verification, temporal hierarchy

Full changelog: v0.2.0...v0.3.0

v0.2.0

Changes

• feat: add Obsidian vault export command
• feat: entity orphan pruning, confidence gating, consolidation tag blocklist, obsidian export
• fix: update digest test to match current summary format
• refactor: remove secondary backup copy from server scheduler
• feat: inject self-model directives into session context
• feat: time-bucketed consolidation and goal-gated retrieval
• feat(dashboard): add Skills page and refactor Graph rendering
• feat: add valence scoring, improve digests, and auto-load condensed skills
• feat(dashboard): merge Search and Timeline into unified Memories page
• feat(dashboard): redesign theme from purple to black/cyan
• fix(ci): add build step, model cache, and mock embeddings for subprocess tests
• fix: correct hash backfill, normalize tag aliases, and harden attribution
• fix(dashboard): restore reliable graph zoom and redraw
• Add lifecycle, contract, concurrency, and UI interaction tests
• feat: harden metadata retrieval, dedupe, and regression lifecycle
• fix: harden server defaults and repair memory/backup integrity
• feat(entities): replace rigid type system with freeform tags
• feat(dashboard): add Graph, Goals, and Memory Links pages
• fix: include FTS matches in candidate pool for keyword search
• feat: add pnpm release script for automated releases

Full changelog: v0.1.0...v0.2.0

v0.1.0 — Initial Release

Exocortex v0.1.0

Personal AI memory system with self-improving retrieval. Local-first, SQLite-backed, MCP-native.

Highlights

Core

• Hybrid retrieval via Reciprocal Rank Fusion (RRF) — fuses vector similarity, full-text search, and graph proximity
• Local embeddings (all-MiniLM-L6-v2) — no external API keys required
• Self-improving scoring — retrieval feedback loop tracks which memories prove useful, adaptive weight tuning adjusts scoring automatically
• Store-time relation discovery — auto-links semantically similar memories at write time
• Multi-hop context expansion — search results include 1-hop linked memories

Memory Intelligence

• Graph-aware search with entity and memory-link proximity boosting
• Health diagnostics (8 automated checks)
• Search friction tracking — logs zero-result queries to identify knowledge gaps
• Consolidation, contradiction detection, and importance recalibration
• Entity graph with centrality analysis, community detection, and co-occurrence densification
• Temporal evolution queries — trace how knowledge about a topic changed over time

Goal System

• Goal tracking with milestones, priority, and deadline support
• Autonomous goal mode — external agent systems can advance goals via MCP tools
• Auto-detection of goal progress from incoming memories

MCP Server (26 tools)

Works with any MCP-compatible client: Claude Code, Codex CLI, Gemini CLI, GitHub Copilot.

memory_store · memory_search · memory_context · memory_get · memory_update · memory_forget · memory_browse · memory_feedback · memory_link · memory_timeline · memory_ingest · memory_digest_session · memory_consolidate · memory_maintenance · memory_graph · memory_entities · memory_ping · memory_decay_preview · goal_create · goal_list · goal_update · goal_log · goal_get · goal_add_milestone · goal_update_milestone · goal_remove_milestone

Dashboard

React + Vite web UI with search, graph visualization, timeline, entity explorer, chat (RAG-powered Q&A), and settings management.

Operations

• Nightly SQLite backup with rotation
• Automatic maintenance pipeline (importance adjustment, archival, entity extraction, trash purge)
• Session-orient hook for Claude Code — auto-loads project context at session start

Setup

pnpm install
pnpm build
node packages/mcp/dist/index.js

See README for full setup instructions.