Add a persistent memory layer that survives repository restarts, server reboots, and team transitions. This fundamentally transforms System Zero from a stateless system to a truly intelligent, learni ng agent.

[MaximilianoAcri]

Problem

Currently, every time System Zero restarts:

• All context is lost
• AI has no memory of past decisions
• Each session starts from zero
• Team members joining mid-project are blind to past context

This is the single biggest limitation of autonomous agents.

Vision

┌─────────────────────────────────────────────────────────────┐
│                    SYSTEM ZERO CORE                        │
├─────────────────────────────────────────────────────────────┤
│  Heartbeat │  Bus  │  Memory  │  LLM  │  Storage │ ???   │
│                                              │         │
│                              ┌───────────────────────────┤ MEMORY │ │
│                              │   PERSISTENT LAYER    │  LYR  │ │
│                              │   ( sobrevive restart ) │       │ │
│                              └──────────────────────│─────────┘         │
├─────────────────────────────────────────────────────────────┤
│                    FILE SYSTEM + DATABASE                │
└─────────────────────────────────────────────────────────────┘

What Gets Persisted

Memory Type	What Survives	Impact
Decisions	Architecture choices, tool selections	High
Bugs	What was fixed, root cause	High
Conventions	Naming, structure, patterns	Medium
User Preferences	How user wants things done	Medium
Discovered Context	Codebase learnings	High

Implementation

Storage Options

Option	Pros	Cons
SQLite	Simple, portable	Single file
PostgreSQL	Robust, scalable	Needs DB
Files (JSON)	Simple, git-trackable	Slower at scale
Vector DB	Semantic search	Complex

Recommended: SQLite for MVP, PostgreSQL for scale

Data Schema

-- Core tables
CREATE TABLE observations (
  id SERIAL PRIMARY KEY,
  title TEXT NOT NULL,
  type TEXT CHECK (type IN (
    'decision', 'architecture', 'bugfix', 
    'pattern', 'config', 'preference'
  )),
  content JSONB NOT NULL,
  project TEXT,
  scope TEXT DEFAULT 'project',
  topic_key TEXT,
  created_at TIMESTAMP DEFAULT NOW(),
  session_id TEXT
);

-- Full-text search
CREATE INDEX obs_search ON observations 
  USING GIN (to_tsvector('spanish', title || ' ' || content));

API

# Save observation (automatic or manual)
sz mem save --type bugfix --title "Fixed N+1 in UserList" \
  --content "Found missing .include(), added eager loading"

# Search memory
sz mem search "N+1"

# Get context for new session
sz mem context --project mi-proyecto

# Topic updates (upsert)
sz mem update --topic architecture/auth-model \
  --content "Switched to JWT from sessions"

Features

1. Automatic Capture

• ✨ Auto-save architecture decisions
• ✨ Auto-save bug fixes with root cause
• ✨ Auto-save team conventions
• ✨ Auto-save user preferences

2. Search

• Full-text search across all memories
• Semantic search (vector similarity)
• Filter by type, project, date

3. Context Recovery

# New team member joins
sz mem context --project misitio --limit 20
# → Returns recent context to get up to speed

4. Integration

• Works with all existing modules
• Compatible with heartbeat, bus, LLM
• No breaking changes

Token Savings

With persistent context:

• Don't repeat questions → ~30% tokens saved
• Reuse shared understanding → ~20% tokens saved
• Smart context loading → ~40% tokens saved
• Total: ~50-70% token reduction

Why This Is EPIC

Before	After
Stateless	Stateful
Starts from zero	Remembers everything
Each tick is new	Builds on past
No team memory	Institutional knowledge
Expensive context	Efficient

---

Comparison with Engram

This proposal is inspired by Engram, a persistent memory system for AI agents with:

• Proactive save triggers
• Semantic search
• Session summaries
• Topic upserts

Would love to collaborate on integrating similar functionality!

---

This would be the most impactful addition to System Zero's core. Game-changing for autonomous AI development.