Hi Product Hunt — I built Talamus because agent context kept disappearing at the worst possible moment: between sessions, tools, and model providers. I wanted memory I could inspect, version, and correct, rather than another opaque hosted black box. The design is deliberately local-first: Markdown remains the source of truth, SQLite/FTS5 makes recall fast, citations point back to exact files, temporal history preserves what was known when, and proposed corrections require review. The quick demo runs with --fake, so evaluating the core does not require an API key or paid model call. I’d especially value feedback from people building coding agents or MCP workflows: which memory failure mode should we test next? The repository includes the runnable demo and committed benchmark artifacts.

Talamus

Durable local-first memory for AI agents, with citations

Durable local-first memory for AI agents, with citations