What actually shipped, from real commit history.
Doubao has no JSON config file — it's a point-and-click custom connector dialog (技能・连接器・伙伴 → 新建 → 新建自定义连接器) supporting both STDIO (for the local server: command + args + env) and HTTP (for the cloud endpoint: URL + custom headers, same x-api-key scheme as Cursor). Verified against a real Doubao client, both paths, including a successful add_memory call on the same account/memory graph used by the other clients.
Added a pluggable embedding provider that talks to any OpenAI-compatible /v1/embeddings server — vLLM, Ollama, llama.cpp, and MLX all expose one as of 2026. Combined with pointing LLM_BASE_URL at the same kind of server, a local install can now avoid cloud API calls entirely, and skip the sentence-transformers/torch dependency when it does. Verified for real against a local Ollama server; the other three engines are documented as architecturally identical but not yet independently tested.
The embedding model was loading synchronously at server startup, blocking it from responding to anything (including a liveness ping) for several seconds. Confirmed against a real container-deploy failure on Glama ("Container exited with code 1 before responding to ping"); the model now loads lazily on first use instead. Also restructured the README around a plain-language quickstart instead of leading with a full dev-environment setup.
`pip install spomory` is now real — published to PyPI, with the local install path added to the official MCP Registry alongside the existing remote entries. A Dockerfile was added so directories like Glama that deploy from a container image can list Spomory too.
Both clients' GUI-based "add a custom MCP server" flows are now verified end-to-end, local and remote alike — Codex's Plugins → Manage → MCPs form (STDIO and Streamable HTTP), and Cursor's Customize → MCPs editor. Also documented the remote connection's real url + headers JSON shape, which hadn't been spelled out before.
Added forget_all_memory after a real user hit the wall forget_memory was deliberately designed around — clearing 4 stored facts meant 4 separate approval-gated calls. It's now one call, and also purges archived relation history so a wipe can't be resurrected later. Also added secret redaction (API keys, tokens, JWTs) before any text reaches the extraction LLM, covering both the authenticated write path and the public /demo/try endpoint. Cursor and Codex CLI are now verified against real installations, not just inferred from official docs.
Added a rate-limited, unauthenticated /demo/try endpoint so anyone can try extraction and retrieval without registering. Retrieval ranking now gives a boost to facts that get restated — and the write path skips the LLM call entirely for low-information filler ("ok", "thanks") that could never contain an extractable fact, saving a real API call.
POST /users/register gained CORS, deduplication, rate limiting, and email verification — the difference between "works in a demo" and "can take public traffic." Relation updates now keep transaction-time history, so a fact's edit history is traceable. Submitted the official MCP Registry's server.json listing the same week.
Local storage is encrypted at rest by default. Retrieval ranking started factoring in how long it's been since a fact was last retrieved, so stale facts naturally lose relevance. A multi-tenant, OAuth 2.1-authenticated cloud MCP server went from scaffolding to verified in one push. Also found and documented a rendering bug that only real English input could expose — the one written up on the blog.
Published the first real LoCoMo benchmark numbers in the README. Caught a Recall@10 regression during the same re-run, tracked down two separate causes, and fixed both — documented, not smoothed over. Added forget_memory, so a user can finally say "forget that I said X" and have it actually happen.
Confirmed add_memory and search_memory both work end-to-end in a real Claude Desktop session — which surfaced a real macOS permissions bug (TCC blocking the MCP subprocess from reading ~/Documents) that got fixed the same day. Grew the RL training set from 27 to 140 real samples and ran a full training pass, after tracking down a reward-function bug that had briefly looked like "not enough data." Also deployed Postgres on a real GPU box that day.
Project scaffolding, pluggable LLM/embedding interfaces, the graph store, and HippoRAG-style retrieval. MCP tool wiring, memory-passport export, and a verified true-delete path. At-rest encryption for local storage. Cloud API auth/quotas, key management, and a deletion audit log. The first real LoCoMo benchmark numbers, and the GRPO training/policy scaffolding that later runs built on.