Integration
Long-term memory for Pochi (TabbyML)
Pochi is TabbyML's agent product (separate repo, `getpochi.com`) — not to be confused with TabbyML/tabby, the self-hosted code-completion engine, which has no agent or tool-calling surface to plug memory into. Pochi has first-class MCP support in its VS Code extension and `pochi` CLI, with config at `~/.pochi/config.jsonc`. The Engram entry is one block in the `mcp` map: `url`, `headers.Authorization`, done.
Install
Three steps: sign up for an Engram API key, paste a BYOK LLM-provider key on /models, then drop the snippet below into Pochi (TabbyML).
Three steps to memory in your agent
- Sign up. Free, no card. You'll land on a Getting Started page that walks the next two steps.
- Add your LLM key. Engram is BYOK. Paste an OpenAI / Anthropic / Groq / Together / Fireworks key and we'll route every extraction and query call through your provider. You pay your provider directly. We never see your inference.
- Paste the snippet below into your agent and restart it. Use
Authorization: Bearer <api-key>with the API key from your portal.
Pochi (TabbyML's agent product) via ~/.pochi/config.jsonc
Not to be confused with TabbyML/tabby (the self-hosted code-completion engine — no agent surface to plug memory into). Both the Pochi CLI and VS Code extension share ~/.pochi/config.jsonc, so one block covers both.
- Add Engram to
~/.pochi/config.jsonc(global) or.pochi/config.jsoncin a project root: - Verify:
pochi mcp listshould printengram ✓ HTTP.
{
"mcp": {
"engram": {
"url": "https://mcp.lumetra.io/mcp/sse",
"headers": {
"Authorization": "Bearer <api-key>"
}
}
}
}Heads up: blocked on Pochi 0.6.8 Anthropic adapter
MCP wiring is correct, but as of 2026-05-20 @getpochi/cli@0.6.8's proxedFetch wrapper references an undefined POCHI_CORS_PROXY_URL_PREFIX, so every Anthropic call crashes with InternalError: Invalid URL. Workarounds: wait for the next Pochi release, switch to a non-Anthropic provider, or patch proxedFetch locally. Tracked at github.com/TabbyML/pochi.
What you can do once memory's wired in
- Add the `engram` block to `~/.pochi/config.jsonc` (global) so every Pochi project sees the six memory tools
- Workspace-scoped install via `.pochi/config.jsonc` in the project root — overrides global, useful for per-project buckets
- Use Pochi's `disabledTools: ["clear_memories"]` on the server entry to keep destructive tools out of the toolbox
- Verify with `pochi mcp list` — `engram` should appear with status ✓ and transport `HTTP`
FAQ
Is this recipe usable today?
The MCP wiring is correct and verified end-to-end, but as of 2026-05-20 the recipe is BLOCKED on upstream Pochi 0.6.8. `@getpochi/cli@0.6.8`'s `proxedFetch` wrapper references an undefined `POCHI_CORS_PROXY_URL_PREFIX` constant, so every Anthropic call fails with `InternalError: Invalid URL` before tools have a chance to run. This is a general Pochi-CLI bug, not specific to MCP or Engram — `--no-mcp` Anthropic chats hit the same crash. Tracked at github.com/TabbyML/pochi.
Are there workarounds?
Three. (1) Wait for the next Pochi release — recommended. (2) Use a non-Anthropic provider in Pochi's config; the bug is in Pochi's Anthropic adapter specifically, OpenAI/etc. paths are likely unaffected. (3) Patch `proxedFetch` in your local `@getpochi/cli` install to fall back to plain `fetch` when the prefix is undefined (one-line change near the top of `dist/cli.js`).
Does this work with the VS Code Pochi extension too?
Yes. Both Pochi surfaces (CLI and VS Code extension) read the same `~/.pochi/config.jsonc`, so one config block covers both. The Anthropic blocker above is CLI-specific; the VS Code extension's Anthropic path may be unaffected, but we've only verified the MCP wiring on the CLI side.
Related integrations
Ship durable memory in Pochi (TabbyML) today
Free tier: 10K memories and 50K retrievals per month. No credit card. Same Engram backend powers all 41 integrations, so memories you write from one client are immediately queryable from the rest.