Show HN: Agent Recall – Open-source, local memory for AI agents (SQLite/MCP)(github.com)

1 pointby MaxNardit3 hours ago2 comments

MaxNardit3 hours ago
Hi HN, I'm Max. I built Agent Recall because AI coding agents forget everything between sessions. Same codebase, same team, same decisions — explained from scratch every morning.
How it works: SQLite-backed knowledge graph with scoped entities, relations, and slots. An MCP server exposes 9 tools so the agent proactively saves facts as you work. At session start, an LLM summarizes the relevant facts into a structured briefing instead of dumping raw data.
What makes it different from a context file: Scope chains with inheritance (same person, different roles per project), bitemporal history (old facts archived, not deleted), and AI briefings that scale beyond what you'd maintain by hand.
Where I need help:
If you use Cursor, Windsurf, or Cline — try the MCP config, tell me what breaks
PRs for other LLM backends (Ollama, local models) welcome
pip install 'agent-recall[mcp]'
Repo: https://github.com/mnardit/agent-recall
jamiecode3 hours ago
SQLite is a good choice for local persistence, but the hard part is usually the retrieval strategy, not the storage. Semantic search gives you relevance but misses recency. Pure recency ordering misses the most important context. The implementations that work best tend to use a weighted combination, with the weights tuned per task type. Curious how Agent Recall handles the cold-start problem when there's no prior context to inform retrieval.
- MaxNardit3 hours ago
  Good observation. Agent Recall doesn't do semantic search at all — deliberately.
  Instead of query → rank → top-k, it loads all entities/slots/observations within the agent's scope chain at session start, then an LLM summarizes them into a structured briefing. Priority is scope relevance (your project > your org > global) and data type (people and active tasks first, historical logs last), with a token budget that truncates lower-priority sections.
  For in-session recall, there's search_nodes — keyword matching, not embeddings. Less powerful but perfectly adequate for structured facts like "who works on project X" or "what did we decide about auth."
  Cold start: first session has no briefing, but the package auto-discovers project files (CLAUDE.md, README.md) and includes them in context, so the agent isn't completely blind. The MCP tools come with proactive-saving instructions, so memory builds organically. After 2-3 sessions the briefing is already useful.
  The tradeoff is explicit: optimized for structured scoped facts (people, decisions, roles), not fuzzy semantic recall. For a coding agent that needs "Alice is the lead on project X, we decided to use REST" — keyword search + scope filtering works. For "find me something vaguely related to that auth discussion" — you'd want embeddings, and that's not what this does.