LLM Memory is a long-lived memory layer for coding agents. It stores reusable decisions, fixes, gotchas, patterns, snippets, scanned reference material, crawled session history, and cross-project insights in PostgreSQL with pgvector so useful context can survive beyond a single conversation.
The retrieval side has grown into more than a note search. It supports project-aware startup context, unacknowledged global alerts, pending capture candidates, semantic and keyword search, contextual and hierarchical retrieval, relationship traversal with graph output, and synthesis flows that help agents turn prior knowledge into a concrete strategy for the current task.
Recent work focused on making the system safer and more operational. Session transcript checkpointing stages reviewable candidates instead of automatically creating durable memories, import/export produces versioned NDJSON backups with collision remapping, destructive forget flows stay behind approval, and the installer now configures Codex or Claude integrations, slash prompts, a Codex skill, Docker-backed Postgres, and user systemd services for the database and local explorer.
The project also includes a visual memory explorer, Ollama/OpenAI/mock embedding providers, historical Claude and Codex session crawling, and a retrieval benchmark harness with golden, scale, and adversarial suites. That gives the tool a regression surface for recall quality instead of relying on subjective impressions of whether memory search feels good.
