Every agent stack has a memory layer. Almost none have a governed one.
The pattern is predictable: grab a vector store, embed text, retrieve by similarity. That works for demos. It breaks when a legal team asks what personal data the agent ingested last Tuesday, or when a compliance audit surfaces a retention violation, or when an agent starts acting on a memory planted by a malicious tool output three sessions ago.
ARES is the architecture that closes those gaps.
What It Is
A governed semantic memory framework for AI agents. ARES organizes agent memory into four typed networks — facts, experiences, beliefs, and observed entities — rather than a single flat vector store. The separation is what makes per-network governance possible: different retention windows, different PII handling, different access policies for different kinds of memory.
The governance stack covers the policies enterprises actually need: automatic PII detection and redaction, configurable retention aligned to GDPR, HIPAA, and SOX, complete audit logging of what was shown to the LLM and when, token budgeting, cryptographic right-to-forget, and prompt injection defense. ARES is framework-agnostic — thin integration shims connect to LangGraph, Pydantic-AI, NVIDIA NIM, and DataRobot without coupling the governance layer to any of them.
ARES 2.0
ARES v1 is pure Python and enforces governance at runtime. ARES 2.0 goes further: the same Python API surface, with a compiled execution layer underneath that makes governance structurally impossible to bypass — not just a runtime check, but a property enforced before the code runs. The v1 regression suite passes against the new backend without modification.
ares-memory-mcp
The MCP bridge layer brings ARES-style memory to Claude Code sessions. Persistent cross-session memory with four categories — FAILED, SOLVED, DECIDED, CONTEXT — designed around how engineering work actually fails. FAILED memories surface first in every search: knowing what not to do is faster than rediscovering what works. This site's own development sessions use it.
Repositories
github.com/skippythemagnificent/ares2 — ARES 2.0
github.com/skippythemagnificent/ares — ARES v1: Python implementation, framework integrations
github.com/skippythemagnificent/ares-memory-mcp — MCP server for Claude Code session memory
github.com/skippythemagnificent/ares-overview — Documentation and strategy materials