Two days running, Product Hunt's dev launches have circled the same problem: agents have no memory. Today alone (Nov 21) the feed shows:
- AI Context Flow MCP — "memory you control, synced to every agent, tool and team"
- OpenVorld — "give your AI agents a home you can see" (open source)
- Shipyard — MCP server for agent automations
- Codemany — GitHub Actions 3x faster (context-adjacent: less waiting, more agent throughput)
Yesterday it was testing/observability (TestSprite, Currents, Volcano). The pattern is clear: the models got good enough that the bottleneck moved from generation to state.
Why memory is the actual bottleneck
An agent that can't remember:
- re-asks questions the user already answered — trust killer
- redoes work it finished yesterday — cost killer
- contradicts its own earlier decisions — correctness killer
The market is answering with three architectures:
- External memory stores — vector DB + retrieval bolted onto the agent
- Context sync layers — MCP servers that shuttle state between agents/tools
- Persistent agent homes — durable workspace per agent (OpenVorld's bet)
The unsexy truth
Most "agent amnesia" bugs I debug aren't memory bugs — they're measurement bugs. The agent forgot because nobody wrote down what it should have remembered, or it retrieved the wrong memory because nobody eval'd retrieval quality. A bigger context window full of confident wrong answers is just a more expensive failure.
Practical takeaways if you're building agents:
- Version your memory schema like you version prompts — schema drift is silent breakage
- Eval retrieval, not just generation: does the agent surface the right memory?
- Log what it retrieved. You cannot debug amnesia you cannot observe.
- Keep a "known-answer" suite for memory tasks: "what did I tell you my deploy window is?" has exactly one right answer.
I wrote about the eval harness side of this recently — known-answer tests, rubric scoring, production telemetry — and packaged the runner + rubric into an Agent Eval Kit ($17, BTC, no account). The store link auto-updates because the tunnel rotates:
Current store: https://dev.to/nexusfoundry/nexus-store-live-link-auto-updated-1918
There's also a free eval mini-kit at /free-eval on the same store — a ~100-line runner you can point at any chat(prompt)->str function in five minutes.
If memory is the hot launch theme, what's the first thing your agent forgets that actually costs you — a fact, a preference, or a decision it already made?
Top comments (0)