AI memory comparison
Tephra vs Letta
Letta is an Apache-2.0 platform for building stateful agents, with a free tier, full self-hosting and memory kept as a git-versioned filesystem. Tephra is a finished hosted memory that the AI apps you connect share. If you are building agents, or you need to run the thing yourself, Letta is the closer match. If you want the AI tools you already use to share one memory without building anything, Tephra is.
Sources read and prices checked 2026-10-10.
Obsidian Metrics makes Tephra and Tracker, the two products this page compares against others. This page is written by the maker of one of the products compared, so read the competitor columns against their own pages, linked below each table. Competitor facts come only from the vendor pages listed under Sources, read on the date shown above; vendors change plans and prices without notice. We are not affiliated with any company named here, and all product names belong to their owners.
Tephra and Letta side by side
| Feature | Tephra | Letta |
|---|---|---|
| What it is | A hosted memory your connected AI apps share: they recall from it and write to it over MCP or REST. | A platform for stateful agents: in its own words, "AI with advanced memory that can learn and self-improve over time". The open-source harness gives an agent a git-versioned memory filesystem, skills, schedules and background subagents that reorganise memory. [1] [2] |
| Entry price | Tephra needs a paid plan: Premium at $19 a month in the Obsidian Metrics Skool community or $29 a month direct, or a $399 one-time seat. No free tier. | A free $0 plan limited to 3 stateful agents, bring-your-own API keys. Pro is $20 a month with up to 20 stateful agents. An API plan is $20 a month with unlimited agents plus $0.10 per active agent per month and $0.00015 per second of tool execution. Teams Pro is $20 per seat. Read 2026-10-10. [3] |
| How usage is limited | No usage metering. There are only abuse limits: 60 requests a minute per key and 120 a minute per member. | Mixed: a cap on the number of stateful agents on personal plans (3 free, 20 on Pro), quota plus pay-as-you-go overage on Letta Auto, per-seat on Teams, and metered credits on developer plans, where credits cover LLM inference and CPU. Tool execution is metered per second. [3] |
| Reach across AI apps | One memory for every AI you connect, led by OAuth sign-in: the Connect page has steps for Claude (web, Desktop, mobile), ChatGPT, VS Code Copilot, Cursor, Claude Code, Gemini CLI, Windsurf, Zed, Mistral Le Chat and Perplexity, plus generic steps for any other MCP app, and a personal key as a fallback, plus REST. AI tools that support neither MCP nor REST calls cannot connect. Tephra's OAuth discovery endpoints were verified live on 2026-10-10. | A hosted remote MCP server at api.letta.com/mcp that any MCP client can reach, and it names Claude Code and Cursor. Authentication is a bearer token only: its docs state plainly that "OAuth-only clients are not supported yet". [4] |
| Seeing and correcting what is stored | A member can view memories, edit or delete them, read edit history, contest a memory, export their own data (app session), and revoke any AI key with an activity log. | Strong. Memory is a git-backed filesystem the agent and you can both inspect and edit, with a memory viewer in the desktop app, a /doctor command that audits placement, duplication and token usage, and a backup taken before memory is restructured. [2] |
| Hosting and self-hosting | Hosted service run by Obsidian Metrics. Self-hosting is not part of the product. We make no encryption-at-rest or compliance-certification claim. | Apache-2.0 on letta-ai/letta and self-hosting is first-class, though the current harness source has moved to letta-ai/letta-code, whose licence we have not verified: its docs describe running Letta fully locally or on your own infrastructure with no Letta account required, with all agent state staying on the device, plus a self-hosted always-on app server. [1] [5] |
| How memories get in | The connected AI writes them by calling a remember tool. Tephra does not capture chats automatically and has no Gmail, Slack or Drive connectors. You can upload text, markdown, JSON, XML and text-layer PDF files. | Automatically, explicitly, and in the background: the agent updates memory when it learns something durable, you can teach it directly, and background subagents review recent conversations to consolidate lessons without interrupting you. It also connects to messaging channels; the repository we read names Slack, Telegram, Discord and custom channels. [2] [1] |
Where Letta is better
- Letta. Is Apache-2.0 on letta-ai/letta - the current harness has moved to letta-ai/letta-code, whose licence we have not verified - and runs fully locally with no account, so it clears a compliance or air-gap review that Tephra cannot. It has a real $0 tier, it writes durable memory without waiting for a model to call a save tool, and background subagents re-read past conversations to consolidate and de-duplicate memory - where Tephra leaves correcting stale facts to you. Its memory is a git repository you can diff and revert in your own editor, and it reaches messaging channels including Slack, Telegram and Discord, which Tephra does not. [1] [3] [2]
Choose Letta if: You are building agents rather than using AI apps, you need to self-host or pass a compliance review, you want a free tier to evaluate on, or you want memory that consolidates and de-duplicates itself instead of asking you to correct it.
Where Tephra is better
- Connected AI apps read and write the same memory directly, rather than moving it between apps by pasting text. Claude's import and export, for example, works by pasted text, and its import path is marked experimental.
- A flat price with no usage metering, instead of credits or request quotas.
- A shared library that Tephra grows every hour from public research and regulatory sources, included with the product. We did not find a competitor that bundles one; that is a qualitative check, not an exhaustive market survey.
- Recall in the app tells you which related memories it did not load and how complete the answer looks.
Choose Tephra if: You want the AI apps you already use to share one memory set up by signing in, and you need it to work in MCP clients that only support OAuth - Letta's hosted MCP is bearer-token only and its docs say OAuth-only clients are not supported yet.
Where Tephra falls short
- Tephra has no free tier and no self-hosted option.
- The MCP tools return ranked records only. The completeness alert is shown in the app, not through MCP.
- No automatic chat capture, and no connectors to Gmail, Slack or Drive.
- No image or audio ingest, and no OCR for scanned PDFs.
- Stale-fact handling is partial.
- No published, vendor-comparable benchmark, and no accuracy lead we can defend.
What we have and have not measured
On 50 held-out LongMemEval-S questions never used in Tephra development, Tephra answered 39 of 50 correctly (78%, 95% confidence interval 65-87%) with a gpt-4o-mini reader, and every evidence session was in its top 10 results for 46 of 46 answerable questions.
Method: LongMemEval-S cleaned, stratified 100-question subset; gpt-4o-mini reader over top-10 retrieved items; official GPT-4o judge prompt; one run, 2026-10-08; held-out questions are the 50 never used in earlier Tephra tuning; competitors run locally from open source under the same harness.
In that same-harness run, none of Tephra's leads over the other systems tested was statistically significant after correcting for multiple comparisons, and its margin over a plain keyword-search baseline was not significant either. We do not claim Tephra is the most accurate memory. We did not benchmark Supermemory, Zep, ChatGPT memory or Claude memory, and we do not compare ourselves with any vendor's self-reported numbers.
Frequently asked
Is Letta free?
Letta lists a free $0 plan limited to 3 stateful agents with bring-your-own API keys, a $20 a month Pro plan with up to 20 agents, and a $20 a month API plan with unlimited agents plus $0.10 per active agent per month and $0.00015 per second of tool execution, as read on 2026-10-10. Tephra has no free tier.
Can I self-host Letta instead of using a hosted memory?
Yes. Letta is Apache-2.0 and its docs describe running it fully locally with no Letta account, with agent state staying on the device. Tephra is cloud-hosted only and cannot be self-hosted, so if self-hosting is a requirement, Letta is the answer and Tephra is not.
Does Letta work in any MCP client?
Letta runs a hosted remote MCP server at api.letta.com/mcp and names Claude Code and Cursor, but authentication is a bearer token: its own documentation says OAuth-only clients are not supported yet. Tephra leads with OAuth sign-in, which is the one reach advantage it holds here.
Did you benchmark Tephra against Letta?
No. We have not benchmarked Letta and we make no accuracy claim against it.
More comparisons
Sources
Competitor facts come only from the pages below, accessed 2026-10-10.
- Letta source repository (Apache-2.0): https://github.com/letta-ai/letta (accessed 2026-10-10)
- Letta documentation, memory: https://docs.letta.com/configuration/memory (accessed 2026-10-10)
- Letta pricing: https://docs.letta.com/pricing (accessed 2026-10-10)
- Letta documentation, hosted MCP: https://docs.letta.com/platform/hosted-mcp (accessed 2026-10-10)
- Letta documentation, self-hosting: https://docs.letta.com/self-hosting (accessed 2026-10-10)