Nine memory labs. Three tiers. Open a tab. Run the code.

Drop a markdown file in your project root. Start Claude Code. Watch it follow the rules. Then prove v2.1.59+ auto-memory survives a session restart with one question.

Drop-in memory layer for any agent. `pip install mem0ai`, set OPENAI_API_KEY, call add() and search(). Vector + graph + reranker behind one managed API.

Spin up an agent on the Letta free tier. Send one message. Then dump Core, Recall, and Archival so you can see which tier holds which fact. The runtime that gave agent memory structure.

Add a fact. Supersede it. Query the graph and watch valid_at / invalid_at write themselves. Graphiti is the canonical graph-beats-vector case when time matters.

Build a 3-node graph, kill it mid-run, resume from the last checkpoint. Then walk history and fork from an earlier snapshot. Not memory, but the layer everyone forgets they need.

One agent, three layers doing different jobs. LangGraph owns the loop, Mem0 owns user facts, a markdown rules file owns behavior. Restart Python twice and watch all three persist correctly.

Implant a false memory, watch your agent take the bait, then ship provenance + allowlist + reconfirm-on-destructive defenses. Don't ship the attack. Ship the guards.

Run the graph engine behind Zep in your own infra. Neo4j in Docker + graphiti-core. Same temporal semantics, full data sovereignty, the tradeoff is you operate Neo4j.

The benchmark every memory blog now cites. We ship the adapter pattern + bench script + per-framework configs for Mem0, Letta, Zep. Run a real bake-off on Monday with a budget.