Precon

Pre-conscious context management for LLM conversations. A multi-tier memory system that automatically retrieves, synthesizes, and persists context across sessions — so your agent remembers what matters without manual bookkeeping.

Features

• Tiered memory hierarchy (L1-L5): active context, SQLite hot storage with FTS5, summaries, pgvector semantic search, and cold archive
• Automatic retrieval: every prompt triggers unconscious retrieval from L2/L3/L4 tiers — no explicit recall needed
• Background persistence: a Haiku agent reviews each turn and decides what to save, trim, or promote
• Semantic search: pgvector-backed L4 tier with 768-dimensional embeddings for meaning-based recall
• Dreaming: configurable cron-based analysis that runs prompts across session history to surface patterns and insights
• Built-in tool system: file operations, code search, diffing, refactoring tools available to the working agent
• Styled REPL: readline-based interactive shell with lipgloss styling and glamour markdown rendering
• Connection checker: precon check validates all external dependencies (SQLite, PostgreSQL, pgvector, embeddings API, Anthropic API)

Prerequisites

• Go 1.25+
• PostgreSQL with pgvector extension (default: localhost:5431)
• Embedding endpoint (RunPod with nomic-embed-text, 768 dimensions)
• Anthropic API key
• just command runner (optional, for build shortcuts)

Installation

# Build from source
go install github.com/hegner123/precon/cmd/precon@latest

# Or clone and build locally
git clone https://github.com/hegner123/precon.git
cd precon
just install

Configuration

Copy the example config to ~/.precon/config.json:

mkdir -p ~/.precon
cp precon.json.example ~/.precon/config.json

Set required environment variables (or add to config file):

export ANTHROPIC_API_KEY="sk-..."
export RUNPOD_API_KEY="..."

Usage

# Start the REPL
precon

# Check all external connections
precon check

# Run dream analysis
precon dream

The REPL automatically manages context across sessions. Prior conversations are retrieved and injected into each prompt based on topic relevance. No manual save/load required.

Architecture

User Prompt -> [Retriever] -> [Synthesizer] -> API Request -> Streaming Response
                   |              |               |
              L2/L3/L4      Compressed ctx    [Persister] -> L2
                                                   |
                                              [Dreamer] -> Reports

Three-layer metaphor:

• Unconscious — autonomic retrieval and persistence on every prompt
• Conscious — the working agent doing actual work with tools
• Dreaming — periodic cross-session analysis and pattern recognition

License

MIT