Every feature, fix, and improvement — from the beginning.
The two largest files are now modular packages. Search is smarter with Knowledge Graph boosting. Vector indexing scales to 10K+ memories. Claims decompose blobs into verifiable atomic facts. Terminals talk to each other automatically.
db.py (2,900 lines) → db/ package with 11 modules. cognitive.py (3,900 lines) → cognitive/ package with 6 modules. All imports unchanged — zero breaking changes.
Memories connected to more Knowledge Graph nodes rank higher. Logarithmic boost based on structural importance bridges semantic and graph retrieval.
Optional hnswlib integration for approximate nearest neighbor search. Auto-activates when memory count exceeds 10,000. Hooks into ingest, search, and GC.
Decompose blob memories into atomic verifiable claims with provenance, contradiction detection, and verification status. Each fact tracked individually.
PostToolUse hook checks inbox automatically via sqlite3. 2-second debounce, skips read-only tools. Zero tokens consumed when no messages waiting.
24 pytest tests covering migrations, CRUD, cosine similarity, KG boost, knowledge graph traversal, and temporal boost. Isolated temp databases per test.
All 9 nightly processes now use Claude CLI (opus) for intelligent decisions instead of Python word-matching. Plus: 3 bugs found by giving GPT-5.4 and Gemini 2.5 full access to the codebase.
Pure Python collects data, then Claude CLI opus makes intelligent decisions. The brain uses its own intelligence to dream — not a calculator.
GPT-5.4 (Codex CLI) + Gemini 2.5 (Gemini CLI) + Claude Opus reviewed the full codebase simultaneously. Found 3 real bugs with exact file:line references.
Retrieved memories are now sanitized to prevent prompt injection. Stored content is treated as quoted evidence, not executable instructions.
The quarantine system now distinguishes confirmation from contradiction. Similar memories that reinforce each other are no longer rejected as conflicts.
NEXO can now evolve its own configuration autonomously through a structured proposal → review → implement cycle. Dual-mode operation ensures safety: auto mode for low-risk changes, review mode requiring explicit approval.
Auto mode applies low-risk improvements (nightly job tuning, threshold adjustments) without intervention. Review mode stages proposals for human approval before execution.
Evolution proposals are generated from session patterns, guard stats, and cognitive metrics. Each proposal includes rationale, risk assessment, and rollback plan.
Proposals are executed via Claude CLI for safe, auditable implementation. Every change is logged and reversible.
Nightly jobs now detect and merge duplicate learnings automatically. Test/debug STM memories are purged during cognitive decay.
NEXO Brain is now licensed under AGPL-3.0. This ensures that any modifications to the cognitive memory system remain open-source when deployed as a service.
The Stop hook no longer says goodbye when a session ends. Farewell messages were a side-effect of the post-mortem enforcement hook, creating confusing output on every session close.
Stop hook no longer says goodbye when session ends. The farewell was an unintended side-effect of the post-mortem enforcement block — removed cleanly without affecting diary write behavior.
The Stop hook now uses decision:block instead of approve to enforce post-mortem execution. This ensures the session diary is always written before Claude exits — no more lost context on abrupt stops.
Stop hook changed from approve to decision:block so post-mortem writing is enforced architecturally, not just suggested.
Fixed infinite block loop triggered when the Stop hook blocked its own post-mortem execution. Added trivial session detection to skip diary write on empty or no-op sessions.
Long sessions no longer lose their thread. NEXO preserves session context across Claude Code compaction events using PreCompact checkpoints and PostCompact re-injection. 8-hour sessions feel like one continuous conversation.
Before compaction, NEXO saves a complete snapshot to SQLite: current task, active files, decisions, errors, reasoning thread, and next step.
After compaction, a structured block is re-injected into context with everything needed to continue seamlessly. Zero "where were we?" moments.
The checkpoint updates automatically every interaction via the heartbeat. No manual action required — continuity is always maintained.
The agent thinks before acting. Architectural inhibitory control validates reasoning quality and physically restricts tools until the plan is sound. Designed via 3-way AI debate (Claude Opus + GPT-5.4 + Gemini 3.1 Pro).
Unknowns detected. Tools restricted to read and search. Agent must gather information before proceeding.
No plan or verification step. Agent can propose but not execute. User reviews before action.
Plan, evidence, and verification present. Full tool access granted. Agent executes with confidence.
v0.10.0 tackles the "stores but forgets" problem. Three new features ensure the right memories surface at the right time — at session boot, during conversation, and when delegating to subagents.
Pre-loads relevant memories at session boot by composing a query from pending followups, due reminders, and last session's topics. No more cold starts.
Bundles all area knowledge (learnings, changes, followups, preferences, cognitive memories) into a single injectable packet for subagent delegation.
Heartbeat detects project keywords in conversation and automatically surfaces relevant learnings. Context arrives proactively — no manual query needed.
Quarantine queue now processes automatically during the nightly cycle. Items are promoted, rejected, or expired based on policy — no manual intervention needed.
Corrections and critical learnings are auto-pinned at ingest. Pinned memories never decay and get a +0.2 boost in search results. User corrections stick permanently.
Dream insights (cross-memory connections) are excluded from default search, reducing noise by 21%. Capped at 50 active insights; excess is archived automatically.
v0.9.0 introduces permanent subconscious memory. Session diaries are now archived forever before cleanup — nothing is ever truly lost. Plus fuzzy credential search for when you can't remember the exact service name.
Session diaries older than 180 days are moved to a permanent diary_archive table instead of being deleted. Search by text, domain, year, or month. nexo_recall automatically falls back to the archive when recent memory has few results.
When exact credential match fails, NEXO searches across service, key, and notes fields. Results are tagged as fuzzy matches so you know they're suggestions, not exact hits.
nexo_recall now has a two-tier search: recent memory first, then diary archive fallback. Your agent never truly forgets — old context surfaces when needed.
v0.8.10 gives your AI agent a relationship graph, a visual brain dashboard, and runs everywhere. 988 bi-temporal nodes with D3 visualization, 6 dashboard pages, and full Linux + Windows support.
Bi-temporal entity-relationship graph with 988 nodes and 896 edges. BFS traversal discovers multi-hop connections. Event-sourced edges with smart dedup (ADD/UPDATE/NOOP). Temporal queries. Interactive D3.js visualization.
Visual interface at localhost:6174 with 6 pages: Overview, Graph, Memory, Somatic, Adaptive, Sessions. FastAPI + D3.js. See your AI brain in real time.
Full Linux and Windows support. Platform-appropriate process managers. Opportunistic maintenance when resources are available.
Generates a full briefing from SQLite: overdue reminders, today's tasks, pending followups, active sessions. Cached for 1 hour.
Mandatory self-critique with 5 questions, session buffer, followup creation, proactive seeds for next session, and reflection trigger.
Saves context before conversation compression. Prevents losing the thread — compaction without a diary means starting from zero.
Captures meaningful tool usage to the Sensory Register. Feeds the Atkinson-Shiffrin memory pipeline silently.
After 3+ sessions, processes the buffer: extracts recurring tasks, error patterns, mood trends. Updates user model. No LLM needed.
Hardwired behavioral foundation: never promise without scheduling, verify before claiming done, audit before delivering, proactive not reactive.
5 configurable axes: autonomy, communication style, honesty level, proactivity, error handling. Set during install, adjustable anytime.
Seamless upgrade from previous versions. Updates hooks, core files, plugins, scripts — never touches your data or customized CLAUDE.md.