Configuration

pgmemory reads configuration from ~/.pgmemory/config.yaml. For solo use, no configuration is needed — embedded PostgreSQL starts automatically with sensible defaults.

Solo use (zero config)

Just run pgmemory start. Embedded PostgreSQL starts on port 7434, the proxy listens on 7432, and everything works.

Team use

The only required setting is the shared PostgreSQL connection string:

postgres_url: "postgres://team-user:password@your-host:5432/pgmemory?sslmode=require"

Or store it securely in your OS keychain:

pgmemory credentials set-postgres-url

The config file will reference it as postgres_url: "keychain:pgmemory/postgres_url".

Full reference

# PostgreSQL connection string. When omitted, embedded PostgreSQL is used.
# postgres_url: "postgres://user:pass@host:5432/pgmemory?sslmode=require"

# Local proxy port (default: 7432)
port: 7432

# Path to the local embedding model (auto-downloaded on first run)
model_path: "~/.pgmemory/models/voyage-4-nano.gguf"

# Embedding dimensions — must match the vector index (default: 1024)
embedding_dim: 1024

# Number of knowledge items retrieved per query (default: 5)
retrieval_top_k: 5

# Maximum tokens of context returned per search (default: 2048)
retrieval_max_tokens: 2048

# Upstream LLM provider URL (default: https://api.anthropic.com)
upstream_anthropic_url: "https://api.anthropic.com"

# Enable LLM-based synthesis for higher quality memory extraction
# Requires an Anthropic API key (set via tray app or credentials command)
llm_synthesis: false

# Quality maintenance settings
steward:
  interval_minutes: 60       # How often the quality sweep runs
  prune_threshold: 0.1       # Score below which low-value items are removed
  grace_period_hours: 24     # Minimum age before removal is considered
  decay_half_days: 90        # How quickly unused knowledge loses relevance
  merge_threshold: 0.88      # Similarity threshold for deduplication
  batch_size: 500            # Items processed per sweep

# Write pipeline noise filtering
pipeline:
  ingest_min_len: 80              # Responses shorter than this skip LLM synthesis
  content_score_pre_gate: 0.35    # Adaptive noise score threshold (0 = disabled)

Tuning

The defaults work well for most use cases. If you need to adjust:

Scenario	What to change
Large team (10+ contributors)	Lower interval_minutes to 30, increase batch_size to 1000
Want to retain knowledge longer	Increase decay_half_days beyond 90
Seeing too many near-duplicates	Lower merge_threshold to 0.85
Want less aggressive cleanup	Lower prune_threshold to 0.05, increase grace_period_hours to 72
Heavy use of ingested sources	Increase batch_size to 2000, lower interval_minutes to 30
Too much noise getting stored	Lower content_score_pre_gate to 0.30 (more aggressive filtering)
Valuable short responses being dropped	Lower ingest_min_len to 40
Want to disable adaptive filtering	Set content_score_pre_gate to 0

Environment variables

The only environment variable relevant to pgmemory is the one that connects your AI tool to the proxy:

export ANTHROPIC_BASE_URL=http://127.0.0.1:7432

This tells Claude Code (or any Anthropic-compatible tool) to route through pgmemory. pgmemory itself reads everything from config.yaml and the OS keychain.