The Straight Dope
AI memory alternatives & pricing · Last updated: July 22, 2026
Most AI tools rent you your own memory. You pay. They keep the keys. Switch models and a big piece of what the system learned about you stays behind.
That is the default. This page is the plain version of the alternative — accurate, sourced where we can, no marketing gloss. This is the anti-marketing page: we tell you what everyone else does, honestly, and link you to their own pages so you can check. The base we are releasing is a portable memory graph you control. Free to self-host. Free hosted option. Open. It works with any model you choose. Internal projects build on the same base. What follows is the honest map of the space so you can feel the difference yourself.
The problem people keep describing
The most common complaint is not "the model is dumb." It is "I have to re-explain the same things again." Domain knowledge, prior decisions, style rules, failed approaches — they evaporate when the session ends or when you change providers. Context windows and RAG help, but they do not own the accumulating model of you.
That tax is real and daily for anyone doing serious work with agents.
Privacy and ownership
If you do not own the memory, you do not own the agent. Model providers and most managed memory layers are structurally incentivized to keep the history inside their systems. Local-first and portable-graph approaches exist because some people prefer the data to travel with them instead of with the vendor.
This is not a feature bullet. It is the structural difference.
The tools — name, what they add, where to check the price
These are the visible options as of mid-2026. No ranking. Just the tools, an honest note on the value they add for a working person, and a link to their own pricing page so you can see the current numbers yourself. Don't take our word for it — go read theirs.
| Tool | What it actually adds | Pricing / page |
|---|---|---|
| Mem0 | Drop-in memory layer with high adoption and many integrations. Good when you want something that works quickly across a lot of frameworks and you're fine with a managed path (self-host is also available). | mem0.ai/pricing |
| Zep / Graphiti | Temporal knowledge graphs. Strong when facts change over time and you need entity/relationship tracking that respects when something was true. More infrastructure than a simple store. | getzep.com/pricing |
| Honcho (Plastic Labs) | Reasoning-first memory with peer modeling. Useful when you want the system to build a deeper running model of a person or agent rather than just storing chunks. Especially visible in the Hermes ecosystem. Background processing means derived insights are not always instant. | honcho.dev (usage-based + startup credits) |
| Local-first layers (Mnemosyne & similar SQLite-native options) |
Zero cloud. Data never leaves the machine unless you choose. Fast, simple, offline. Less multi-peer modeling and team features so far, but the ownership is absolute. | Open source / free — search GitHub for current projects |
| multiprovider.cloud (this base) | Portable graph you own. Works with any model. Free self-host or free hosted option. Open for inspection. The same foundation internal projects use. No rental. | multiprovider.cloud — free |
Prices and limits change. Always check the linked pages for the current numbers. The links above go straight to their official pricing or product pages. This page is dated; see the changelog below for when it was last checked.
What this base is
You own the memory. It is a graph on a disk you control.
Free. Self-host or use the free hosted option.
Portable. Route to any model. History travels with you.
Open. Read the code. Run it. Change it.
Data only moves when you explicitly share it.
That is the base being released at disability:IN (July 26–31, Dallas). Internal work adds layers on top of the same foundation. The public release is not a cut-down teaser.
Known gaps in this draft
We keep this honest by naming what's still thin. This page is maintained; these are the next research passes:
- Exact complaint-volume numbers are still being measured. Themes are consistent; hard counts need a systematic pass.
- Independent side-by-side latency and token numbers (not vendor papers) are incomplete.
- Full privacy policy and data-handling details for our hosted option are being expanded and linked.
- Concrete self-host cost estimates and common failure modes for this base are not yet expanded.
- disability:IN-specific framing (why owned memory matters in accessibility workflows) is still open.
- Newer pure-local projects beyond the ones named need another scan.
Changelog
- 2026-07-22 — First publication. Initial tool map (Mem0, Zep/Graphiti, Honcho, local-first, multiprovider.cloud), pricing links, known-gaps section. Sources: public GitHub signals + community discourse, mid-2026.
Spot something out of date or wrong? Tell us at privacy@hplcompany.com — we'll fix it and note it here. That's the deal.
Supported by The Cottonwood Collection and The HPL Company · Privacy · Terms