Show HN: Supe – Give your AI agent a brain, not just memory (opens in new tab)

I built Supe because I wanted Claude to analyze game binaries but not modify them. The constraints grew into something more general: a cognitive architecture for AI agents.

The problem: Most agent frameworks treat memory as flat storage. Store a key, get a value. That's not how useful memory works.

What Supe provides:

1. Neural Memory - Hebbian learning ("fire together, wire together"). Cards connected by synaptic links that strengthen with co-activation and decay with disuse. Spreading activation for recall. Hubs emerge naturally.

2. Validation Gates - Python functions that run before/after tool executions. Block `rm -rf`, enforce read-only mode, whitelist commands. Code, not configuration.

3. Proof-of-Work - SHA256 hashes chain every execution. Tamper with logs and proofs won't verify.

4. Cognitive Hierarchy - Moments (sessions) → Cards (knowledge units) → Buffers (raw data). Not flat.

5. Semantic Relations - 7 typed connections: CAUSES, IMPLIES, CONTRADICTS, SUPPORTS, DEPENDS_ON, EQUALS, TRANSFORMS.

Example gate:

  @agent.register_gate("safe")
  def safe(record, phase) -> GateResult:
      if "rm -rf" in record.tool_input.get("command", ""):
          return GateResult("safe", False, "BLOCKED")
      return GateResult("safe", True, "OK")

Example neural recall:

  neural.add_card(1, {"title": "OAuth"})
  neural.add_card(2, {"title": "Login"})
  neural.connect(1, 2)  # Hebbian learning
  results = neural.recall("authentication")  # Spreading activation

343 tests. MIT license. Works with Claude SDK.

pip install supe