Stem by Awase MCP server
Commit Oct 26, 2026Provides Node.js and Python clients to enable persistent AI memory and context sharing across different assistant platforms like ChatGPT, Claude, and Cursor.
Also known as devawasedev/stem-agent-cookbook
- Stars
- 0
- Tools
- 5
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Oct 26, 2026
- License
- MIT
What is Stem by Awase MCP?
Stem by Awase provides Node.js and Python clients to enable persistent AI memory and context sharing across different assistant platforms like ChatGPT, Claude, and Cursor. It exposes 5 tools, such as remember, recall, graph.
Summary generated from the project's README
Stem by Awase MCP tools 5
Detected from the READMErememberrecallgraphlinkforget
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachabledevawasedev/stem-agent-cookbook is public and reachablegithub.com/devawasedev/stem-agent-cookbook ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-10-26github.com/devawasedev/stem-agent-cookbook ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model5 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNo tool named like delete, drop or execPass
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubSTEM agent cookbook — Stem by Awase
Persistent AI memory across ChatGPT, Claude and Cursor: save selected context, then recall it through an authorized connection to the same STEM account. See the agent hub for supported hosts and connection guides.
Connect to https://awase.dev/stem, save a selected piece of context and retrieve it from another compatible assistant connected to the same STEM account. This directory is a standalone, dependency-free starting point for Node.js and Python clients. It contains no account credentials or production memories.
Public entry: STEM developers. Connection guide: STEM cookbook. Product capability reference: STEM for agents.
Dated evidence: selected-context handoff
On October 25, 2026, an isolated synthetic test used ChatGPT web, Claude web and Cursor Agents. ChatGPT reported saving a note; expanded native Claude and Cursor recall responses returned its same reported memory ID and exact text. ChatGPT's save and deletion payloads were not independently captured. Cleanup was assistant-reported deletion plus zero results in one bounded Claude recall, not an account-wide erasure audit.
Cursor used User MCP, This Mac, with no repository. This establishes one observed recall through that configuration; it does not verify saving in Cursor, the Cursor plugin's runtime, every host/version, full-chat import or ongoing transcript synchronization. Read the dated native demonstration and limitations and compatibility evidence.
Run locally first
Requirements: Node.js 22+ and Python 3.9+. No npm install or pip install is needed for the core examples. Unzip the download and enter stem-agent-cookbook/.
node --test test/*.test.mjs
The tests run a disposable server bound to 127.0.0.1 on a random port. They exercise both languages, synthetic OAuth PKCE/DCR, cleanup, credential redaction, redirect refusal, and scorer arithmetic. They do not contact an AI provider or your STEM account. The source-catalog parity test runs in the STEM repository and skips in the standalone archive because the production source is not bundled.
For a visible walkthrough, start the fixture in one terminal:
node fixture-server.mjs
Copy the printed URL into a second terminal (replace PORT):
node smoke.mjs --fixture http://127.0.0.1:PORT/stem
python3 smoke.py --fixture http://127.0.0.1:PORT/stem
Each example saves one uniquely marked synthetic note, recalls that exact note, inspects the graph, and deletes only the numeric ID returned by its own save. It then checks that the ID does not appear in another bounded recall. It never deletes IDs merely found by a search. Stop the fixture with Ctrl-C; its in-memory data disappears.
The fixture approximates the public wire shapes; its exact-substring search, simple graph and automatic OAuth grant are not STEM production retrieval, authentication, account isolation, consent enforcement, TLS or provider compatibility tests. All data is synthetic. Do not expose this fixture to a network or use it as a backend.
Reproduce a Node-to-Python shared-memory demo
DEMO.md walks through one Node save followed by an exact recall from a fresh Python MCP session, with exact-ID cleanup and a sanitized timing receipt. Both clients use one user-authorized OAuth grant. This demonstrates two language clients sharing saved context, not two independently authorized LLM hosts.
node shared_memory_demo.mjs --fixture http://127.0.0.1:PORT/stem
# Only after choosing a dedicated synthetic account:
node shared_memory_demo.mjs --live-synthetic --receipt /absolute/path/to/new-receipt.json
The child receives credentials over stdin rather than command arguments, files or environment variables. Receipt fields distinguish local fixtures, live attempts, confirmed retrieval and cleanup. The October 25 live result includes a sanitized receipt with actual timings and exact-source hashes; it is a developer-client run, not an LLM-host test or screen recording.
Explicit live synthetic smoke
Only run this when you intend to authorize a real test connection. Prefer a dedicated STEM account containing synthetic data. These commands create an OAuth client registration, ask you to sign in in your browser, save one synthetic memory to that account, remove that exact memory, and attempt token revocation:
node smoke.mjs --live-synthetic
# Or use Python instead, not both unless you want two registrations:
python3 smoke.py --live-synthetic
Select manual saving for this connection in the first-party authorization screen. The examples request only stem, and they never call remember_automatically or change memory settings. They do not reuse existing browser tokens, account keys, environment secrets or connector credentials. A normal browser sign-in is still needed. The authorization URL printed by the CLI contains a public client ID, state and PKCE challenge; it never contains a token, code verifier or authorization code. Callback logging is disabled. Tokens stay in process memory, are sent only in the Authorization header to the configured resource, and are never saved or printed. Revocation attempts cover the returned access/refresh tokens. A registered OAuth client may remain visible after the smoke; manage that specific test connection in STEM settings if desired.
Failures are not silently retried: remember and link can duplicate writes. If a response is lost before a new ID is known, the client cannot safely infer what to delete; inspect the uniquely marked synthetic note yourself. If a known-ID cleanup fails, the CLI reports that ID. A deletion acknowledgment plus absence in bounded recall is useful evidence, not an account-wide erasure audit. The graph is also bounded, so a missing graph node is reported rather than treated as proof of data loss. The examples do not log recalled text, but using a personal account may expose other ranked candidate memories to the client process. Use a clean test account to avoid that exposure.
How authorization works
oauth.mjs and oauth_client.py discover STEM's protected-resource metadata at /.well-known/oauth-protected-resource/stem, verify the exact resource, and pin the authorization server and its endpoints to the resource's origin. They require public-client authentication and S256 PKCE, register a loopback callback using DCR, generate fresh state/verifier values, validate state (and issuer when returned), and exchange the one-use code with the same redirect URI and resource=https://awase.dev/stem. The client rejects unexpected granted scopes. TLS validation remains enabled; redirects are refused for metadata/token/MCP requests so credentials cannot follow another origin.
This is a deliberately STEM-specific tutorial, not a generic discovery library. It uses the supported legacy 2025-06-18 JSON-response Streamable HTTP subset and does not implement SSE subscriptions, refresh rotation, pagination, the 2026-07-28 request model, CIMD, or every OAuth provider. It rejects an unsupported negotiated protocol rather than pretending to support it. A server's structuredContent is used when present, with JSON text fallback. For a full application use a maintained MCP SDK, preserve per-user token boundaries, and test your host/version. DCR remains a backward-compatibility mechanism in the current MCP authorization specification; a production client should use CIMD when both sides support it. Legacy authorization requirements explain the PKCE, discovery and resource parameters used here.
Python requests identify the actual cookbook and Python runtime with a stem-agent-cookbook/1.0.0 Python/... User-Agent. They do not impersonate a browser or search crawler. A blocked or challenged request still fails; the client does not disable TLS, rotate identities, solve challenges or change server protections.
Reuse project context with a skill
For an assistant that supports user-installed skills, review the portable STEM project-context skill. It recalls relevant decisions before work and saves only selected, approved context. Installing instructions does not connect an account or grant tool access; use the host setup guides and review your host’s installation flow. The Cursor plugin source packages its Cursor-specific version. Neither version imports full conversations or enables automatic saving.
Actual core tools
fixtures/tool-catalog.json is a checked snapshot of STEM's five core tool definitions. Fetch tools/list on your actual connection to discover its current catalog. Some connections additionally expose consent-gated tools; their availability is not assumed by the smoke.
| Tool | Example arguments | What matters |
|---|---|---|
remember |
{"text":"Selected context","source":"notes","type":"note","autolink":false} |
Explicit save; returns numeric ID; repeated calls create duplicates. |
recall |
{"query":"Project Lantern","limit":8} |
Ranked, bounded candidates; inspect relevance; does not read unsaved chats. |
graph |
{"limit":1000} |
Short labels and display IDs such as m12; not full text or total account inventory. |
link |
{"from":12,"to":11,"type":"builds_on"} |
Numeric IDs from this account; creates a relationship between memories, not assistant accounts. |
forget |
{"id":12} |
Destructive; delete only the item the user identified; no undo. |
An origin label such as claude or chatgpt is metadata, not authorization to fetch that provider's chat history. Selected context must be supplied by the user or already saved in STEM. Connecting the same account in two supported hosts enables shared saved context; it does not give STEM universal transcript access. Host confirmation requirements and explicit no-save instructions still apply.
Framework recipes and evaluations
recipes/README.md supplies small FastMCP and LangChain recipes based on current official documentation. They are optional integrations with their own dependencies; they are not included in the dependency-free runtime claim. No provider calls or framework compatibility scores are reported from merely reviewing their code.
evals/golden-prompts.json defines 16 synthetic direct, indirect, negative and consent tasks. evals/evaluate.mjs grades supplied recorded results, with separate tool-selection precision/recall, incorrect saves, selected-memory retrieval, latency and token fields. Missing measurements stay missing. See evals/README.md for the data contract, fixture-ID mapping and limitations.
node evals/evaluate.mjs /path/to/synthetic-recorded-results.json
node record_baseline.mjs
evals/baseline.json contains only the measured local deterministic test summary, source hashes and explicit null model metrics. The scorer's hand-authored arithmetic tests are not a model benchmark. A live OAuth smoke is still not an LLM tool-selection evaluation. Publish model/provider/harness versions, coverage, repeated trials and observed outcomes with any future model result.
Rebuild the download
python3 build_archive.py
# Optional explicit output path:
python3 build_archive.py --output /absolute/output/stem-agent-cookbook.zip
The builder uses a fixed file allowlist, rejects symlink inputs and normalizes ZIP metadata. It includes a SHA256SUMS manifest and prints the ZIP SHA-256. It excludes tokens, environment files, unallowlisted local result recordings, Python caches, dependency folders and generated archives. The explicit allowlist includes one curated, sanitized live-demo receipt. Default output is dist/stem-agent-cookbook.zip; archive bytes are reproducible for the same source files and compression implementation. The output archive is a distribution artifact, not a credential store.
License scope
The example code and documentation in this cookbook are available under the MIT License, copyright 2026 Awase LLC. This license applies only to the cookbook directory and its distributed copies. It does not license STEM's backend, account data, hosted service, branding or other repository directories. Service access remains subject to its published terms and each user's authorization.
Stem by Awase MCP questions
What tools does Stem by Awase MCP provide?
It exposes 5 tools, including remember, recall, graph, link, forget.
What does Stem allow me to do with my AI assistants?
It enables persistent AI memory across platforms like ChatGPT, Claude, and Cursor by allowing you to save selected context and recall it through an authorized connection to the same STEM account.
Does connecting Stem give it access to my full chat history?
No, connecting the same account in different hosts enables shared saved context, but it does not grant universal transcript access or the ability to read unsaved chats.