How PA·dev compares
If you're building agents, you've probably already reached for one of these. Most of them solve one slice well — a vector store, a memory module, a stateful agent, a memory OS. PA·dev is the substrate underneath all of those, plus the planner that runs the work. Here is the honest version, tool by tool.
vs Pinecone
A managed vector database. You run the embeddings and store the vectors as a piece of infrastructure you operate.
- Your brain is Markdown in a repo you own, not rows in someone else's database.
- When a zone gets big, the index is a local file beside the Markdown — regenerated on boot, with no service to provision or pay for.
- Leave any time and take the folder with you. There is nothing to export, because you already have it.
vs LangChain
A framework. You wire a memory module to a separate vector store and hold the two of them together in your own code.
- Memory, personas, and privacy zones are one primitive here, not three libraries you integrate and keep in sync.
- A zone is enforced server-side at the row, before the model ever sees the data — not a prompt instruction you hope holds.
- It's hosted and reachable over REST and MCP, so the same brain answers your terminal and your production cron job.
vs Letta
Stateful agents with self-editing memory, scoped to each agent process while it runs.
- One brain, many clients: Claude Code, Cursor, an MCP client, and a cron job all read and write the same layer.
- State isn't trapped inside one running agent — it's a substrate that every agent shares.
- Personas are reusable specs, so a team shares one agent instead of each person rebuilding their own.
vs Mem0
A memory API for AI apps — store facts and retrieve them over time.
- It's not only recall. The scaffolds endpoint plans the work and dispatches sub-agents to actually do it.
- Personas and privacy zones ship in the same substrate, not as a separate add-on you bolt on later.
- Bring your own model and your own backend — a GitHub repo or a local folder, six providers through one dispatcher.
vs EverOS
Best-in-class for self-hosted memory benchmarks — 93% on LoCoMo. Apache-2.0, open-source, local-first, with a research paper behind it. The wedge is recall: how well an agent remembers.
- Different category, same audience: EverOS optimizes how an agent remembers; PA·dev is the layer that makes the agent act on what it remembers — scaffolding, dispatcher, connector write-actions, approval gates.
- Memory is one primitive here, sitting next to personas and privacy zones — not the whole product.
- Hosted and multi-tenant on one key, with BYO model, instead of a runtime you self-host and operate.
- If you're benchmark-chasing pure recall, EverOS is the call. If your agent needs to act inside a customer's tools, PA·dev. Here's the full, fair breakdown.
Built by Whited Consulting — used in production at Tennessee Valley Exteriors, Patrick @ Fresh Page, and Buildout Studios client builds.
The fastest way to judge it is to make a call.
The MCP server is open source, so you can run it locally and mock the API before you pay for anything. The quickstart gets you there.