Memory System

Gee-Code has a three-layer memory system that persists across sessions. The AI learns your preferences, remembers project context, and builds on past conversations — so you never have to repeat yourself.

Three Layers

Layer 1: Resources (Raw History)

Raw conversation transcripts from past sessions. Searchable but unprocessed — this is the complete record of everything that was said and done.

Layer 2: Items (Extracted Facts)

Atomic facts extracted from conversations. These are specific pieces of information:

• “Prefers Python over JavaScript”
• “Works at Acme Corp on the billing system”
• “Uses VS Code with vim keybindings”

Facts are automatically extracted after each session. They’re matched per-turn to inject relevant context into the conversation.

Layer 3: Categories (Evolved Summaries)

High-level summaries across 7 categories, refined and evolved over time:

Category	Examples
personal_info	Name, location, devices, family
coding_preferences	Languages, frameworks, style rules
projects	Active projects, architecture decisions
work_context	Team, role, workflows
tools_and_setup	IDE, OS, shell, terminal configuration
communication_style	Preferred level of detail, tone
general	Miscellaneous facts

These summaries are the richest view of what Gee-Code knows about you. They evolve as new facts are extracted and patterns emerge.

How It Works

Each turn, the memory system:

1. Searches Layer 2 and 3 for facts relevant to the current conversation
2. Injects matching memories into the AI’s context
3. After sessions, new facts are automatically extracted and stored

You’ll see memories appear as <long-term-memory> blocks in the AI’s context. The AI references these naturally — applying your preferences without asking, connecting current requests to past work, and anticipating needs based on patterns.

Memory Tools

The AI can interact with memory explicitly:

Remember — Store a fact

Remember("Prefers functional components over class components", category="coding_preferences")

The AI does this proactively when you share preferences, make decisions, or reveal context worth preserving.

RecallMemory — Search memory

RecallMemory(query="database preferences")
RecallMemory()    # Full overview of all memories

The AI searches memory when stuck, working in an unfamiliar area, or when you reference something from a past session.

ForgetMemory — Delete a fact

ForgetMemory(memory_id="abc123")

Remove outdated or incorrect information.

Proactive vs. Reactive

• Proactive: The AI stores important facts as it encounters them — preferences revealed through feedback, architecture decisions, personal information you share
• Reactive: The AI searches memory when past context might help — when stuck, starting a new task, or when you say “remember when we…”

What Gets Remembered

Stored automatically:

• Personal information shared by you
• Coding preferences revealed through feedback
• Project architecture decisions
• Tool and workflow preferences
• Communication style patterns

Not stored:

• Temporary debugging details
• One-off task specifics
• Information only relevant to the current session

Procedural Patterns

Over time, the system builds behavioral patterns — profiles of how you work:

“When debugging, searches with grep first, then reads files, then applies fixes”

These help the AI match your working style rather than imposing its own approach.

Memory Storage

Memory is stored locally:

• Global memory: ~/.gee-code/memory/
• Mode-specific memory: ~/.gee-code/memory/modes/{mode-name}/

Each Gee (autonomous mode) gets its own memory space, so their learned context doesn’t bleed into each other.

Next Steps

• Sessions & Continuity — session lifecycle and ledger
• Configuration — customize memory behavior
• Gees & Teams — autonomous agents with their own memory