Thread-safe SQLite with WAL mode and connection pooling, passing 100% of concurrent stress tests. Built on formal schema migrations and trust auto-detection, on top of 21 cognitive tools implementing cognitive science concepts that go beyond basic memory storage.
Only novel information is stored. Redundant content that matches existing memories is rejected at write time, keeping memory clean without manual curation.
4-layer defense against memory poisoning: injection detection, encoding analysis, behavioral anomaly scoring, and credential scanning.
New facts enter quarantine status and must pass a promotion policy before becoming trusted knowledge. Prevents unverified information from influencing decisions.
Auto-detects and redacts API keys, tokens, passwords, and other sensitive data before storage. Secrets never reach the vector database.
Granular lifecycle states. Pin = never decays (critical knowledge). Snooze = temporarily hidden. Archive = cold storage (searchable but inactive).
Batch cosine deduplication during the 03:00 sleep cycle. Respects sibling discrimination — similar memories about different contexts are kept separate.
Discovers hidden connections between recent memories during sleep. Surfaces non-obvious patterns like "these three bugs all relate to the same root cause."
Generates hypothetical answer embeddings for richer semantic search. Instead of searching for your query, it imagines what a helpful memory would look like, then searches for that.
Graph-based co-activation network. Memories retrieved together reinforce each other's connections, building an associative web that improves over time.
Transparent score breakdown for every retrieval result. Shows exactly why a memory was returned: semantic similarity, recency, access frequency, and co-activation bonuses.
Context-triggered reminders that fire when conversation topics match, not just by date. "Remind me about X when we discuss Y" works naturally.
Extracts decisions, corrections, and factual statements from conversations automatically. You don't need to say "remember this" — the system detects what's worth storing.
Structured migration system with schema_migrations table tracking all database changes. Safe, reversible schema evolution for production systems. Prevents data loss during upgrades.
nexo_heartbeat now auto-detects trust events (thanks, corrections, delegation, repeated errors) directly from user text. No manual trust calls needed. Deltas configurable via DB table.
WAL mode + connection pooling enables fully concurrent multi-agent access. 100% pass rate on concurrent stress tests. Multiple Claude instances can read and write simultaneously without locking or data corruption.
Session diaries are written incrementally during the session, not just at close. Orphan sessions (crashed or abandoned) are auto-detected and closed with a synthetic summary, so context is never lost even when the agent dies mid-task.
Memories connected to more Knowledge Graph nodes rank higher. Logarithmic boost (capped at +0.08) bridges semantic and structural retrieval, so structurally important memories surface first.
Optional hnswlib integration for approximate nearest neighbor search. Auto-activates at 10,000+ memories. Hooks into ingest, search, and garbage collection. Graceful fallback to brute-force.
Decompose blob memories into atomic verifiable claims. Each claim has provenance, confidence score, and verification status. Contradiction detection across sources ensures factual consistency.
Open source, AGPL-3.0 licensed, and built for builders who want their AI to actually remember.