Grounded recall you can prove. Simulation you can fence. mem20 separates real memory from imagined memory at the storage layer — and enforces that separation with a contamination firewall.
Today's AI agents hallucinate, confabulate, and forget — not because they're broken, but because their memory systems treat every stored token as equally real. mem20 fixes that by partitioning memory at the engine level:
Facts derived from real observations, user statements, or events. Indexed, searchable, and always attributable.
Hypotheticals, counterfactuals, and imagined scenarios — fenced off from recall indexes, never contaminating grounded truth.
Simulated records become grounded only when paired with real-world prediction resolution — not free-text self-assertion.
Run memory_audit_contamination() anytime. A clean system reports rate 0.0 — and the firewall makes that invariant structural.
Speaks Model Context Protocol over stdio. Drop into Hermes, Claude, Cursor, or any MCP host. 97 tools, 13 domains.
Blender and Unity ship as optional modules. Add Figma, Unreal, or anything else via the blank integration template.
mem20 is not a wrapper around a vector DB. It's a purpose-built memory engine with protocol, cognition, storage, and governance layers — all enforcing a single rule: simulated content never reaches grounded indexes.
Vector and BM25 indexes are grounded-only by construction. _add_to_vector_index() and _add_to_bm25_index() call _assert_grounded() on the lowest write path. Contamination is impossible regardless of which tool invoked the write.
| Layer | What it does | Where it lives |
|---|---|---|
| Transport | MCP server, JSON-RPC 2.0 over stdio | mcp/mcp_server.py |
| Cognition | Plan, chain, reason, reflect, working memory | cog/cognitive_engine.py |
| Memory | Grounded vs simulated store, vector + BM25 indexes | memory_engine/memory.py |
| Governance | Contamination guards, epistemic veto, audit | Enforced at the store engine level |
126 bots, isolated namespaces, cross-contamination rate 0%. Persistent per-bot memory that survives restarts and model swaps.
Working memory that persists across sessions. An agent that remembers what it was doing yesterday, last week, last month.
Run counterfactuals without polluting grounded recall. Promote simulations to facts only when evidence supports it.
Join the beta. Get early access, direct input on the roadmap, and a channel to the team building it.
Sign Up for Beta →