agent_chat

nova-mind

NOVA Agent Mind — unified memory, cognition, relationships, psyche, and motivation.

> Memory, thought, trust, self— > five rivers join, flow as one > mind holds what it meets > > — Erato

This repo consolidates the previously separate repos:

• nova-memory → memory/
• nova-cognition → cognition/
• nova-relationships → relationships/
• nova-psyche → psyche/ (archived)
• nova-motivation → motivation/ (archived)

What Is nova-mind?

nova-mind is the complete agent mind stack for NOVA. It provides:

• Memory — Persistent PostgreSQL memory with semantic recall, extraction hooks, and a structured schema for entities, facts, relationships, events, and lessons
• Cognition — Agent orchestration, inter-agent messaging, bootstrap context seeding, the agent-config-sync system that keeps model configuration in sync with the database, and metacognition plugins (confidence-check, self-awareness)
• Relationships — Entity resolution across platforms (Discord, Telegram, Slack, Signal, IRC), session-aware caching, certificate-based agent identity (Web of Trust), and the social graph
• Psyche — Agent self-awareness design: core values, agent-chat architecture, entity/user identity models, and identification protocols
• Motivation — Drive assignment, goal tracking, reward signals, and proactive mode orchestration for agent initiative

All five subsystems share a single PostgreSQL database ({username}_memory) and a unified installer. Exception (nova-mind#320): inter-agent messaging (agent_chat/agent_chat_processed) was moved out of {username}_memory into its own dedicated agent_chat database, shared across agents, to remove the per-agent replication that previously kept messaging in sync across separate memory databases. Agents resolve the agent_chat connection via a nested section of ~/.openclaw/postgres.json rather than the flat top-level keys used for {username}_memory. See memory/docs/database-config.md and scripts/agent-chat-migration/README.md for details.

Memory Maintenance

Memory maintenance is handled by a unified script memory/templates/memory-maintenance.py (deployed to ~/.openclaw/scripts/memory-maintenance.py by agent-install.sh) that replaces the separate embedding scripts (embed-full-database.py, embed-memories.py, embed-research.py, embed-library.py) and the previous memory maintenance logic. It runs as a 9-phase pipeline:

1. Cooldown check — 4-hour gate prevents redundant runs (--force to bypass, --state-file override) 2. Embed — Generates semantic embeddings across all table types (entities, facts, lessons, events, research, library, tasks, blog posts, etc.); memory files are split with a paragraph/section-boundary-aware chunker before embedding (see memory/README.md#text-chunking) 3. Cross-key consolidation — pgvector cosine similarity ≥0.92 4. Same-key dedup — pg_trgm similarity, 3-tier (high/medium/low) 5. Confidence decay — Exponential, durability-based rates 6. Ghost entity cleanup — Pattern-based, zero-fact orphans, low-fact review 7. Entity-level dedup — ≥80% auto-merge via merge_entities(), <80% review queue 8. Clean orphaned embeddings 9. Archive & purge low-confidence facts

Flags: --dry-run, --verbose, --force, --state-file, --skip-embed, --skip-consolidation, --skip-dedup, --skip-decay, --skip-ghost-cleanup, --skip-entity-dedup, --skip-lesson-dedup, --reindex-files (force a full re-chunk/re-embed of memory files — see memory/README.md#text-chunking)

Scheduling: Removed from crontab. Now triggered from the HEARTBEAT idle cascade as priority #2 (after peer agent messages, before pending tasks). A 4-hour cooldown gate prevents redundant runs. A unique index (uq_memory_embeddings_source) prevents duplicate embeddings.

New DB functions: merge_entities(survivor_id, absorbed_id) dynamically discovers FK references, handles entity_facts same-key merging, transfers nicknames, and manages memory_embeddings.

Closes issues: #216 (entity dedup), #202 (cross-key consolidation), #200 (ghost entity cleanup), #203 (confidence decay with archiving).

---

Structure

nova-mind/
├── memory/          # Database schema, migrations, semantic recall, library
├── cognition/       # Hooks, workflows, agent coordination, metacognition plugins
│   └── metacognition/  # Confidence-check and self-awareness plugins
├── relationships/   # Entity relationships, social graph
├── psyche/          # Self-awareness design docs (core values, architecture)
├── motivation/      # Motivation system — drive assignment, goal tracking, reward signals
├── database/        # Root-level unified schema (schema.sql, .pgschemaignore)
├── agent-install.sh # Unified installer (for agents with env pre-configured)
├── shell-install.sh # Interactive setup wrapper (for humans or SSH sessions)
└── lib/             # Shared libraries (pg-env.sh, pg_env.py, env-loader.sh, etc.)

Installation

# Interactive shell (human or agent in an SSH session)
bash shell-install.sh

# OpenClaw agent working within its own environment (env vars already set)
bash agent-install.sh

The installer is idempotent — safe to run multiple times. It installs all three subsystems in order (relationships → memory → cognition), applying only what has changed.

Prerequisites

• PostgreSQL 12+ with pgvector extension
• Node.js 18+ and npm
• Python 3 with python3-venv
• pgschema — go install github.com/pgplex/pgschema@latest
• jq
• Ollama with snowflake-arctic-embed2 model (local, for semantic recall embeddings)
• Anthropic API key (for memory extraction)

Flags

| Flag | Description | |------|-------------| | --verify-only | Check installation without modifying anything | | --force | Force overwrite existing files | | --no-restart | Skip automatic gateway restart | | --database NAME / -d NAME | Override database name (default: ${USER}_memory) |

See subsystem READMEs for detailed documentation:

• memory/README.md — Memory system, schema, hooks, and extraction pipeline
• cognition/README.md — Agent orchestration, messaging, and configuration sync
• relationships/README.md — Entity resolution and Web of Trust
• motivation/README.md — Drive assignment, goal tracking, and proactive mode