Repository navigation
Migrate Agent Framework integration to Supermemory v5 - #1786
Conversation
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Automations to automatically generate PRs for you. |
Deploying with
|
| Status | Name | Latest Commit | Updated (UTC) |
|---|---|---|---|
| ✅ Deployment successful! View logs |
supermemory-mcp | 6917c16 | Oct 08 2026, 11:17 PM |
Deploying with
|
| Status | Name | Latest Commit | Preview URL | Updated (UTC) |
|---|---|---|---|---|
| ✅ Deployment successful! View logs |
supermemory-app | 6917c16 | Commit Preview URL Branch Preview URL |
Oct 08 2026, 11:17 PM |
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
|
Note Production impact unlikely. No production resources are plausibly affected by this change. Dependency changes
Polylane analysed Rate this review: 👍 helpful · 👎 not helpful · 😕 confusing |
There was a problem hiding this comment.
Approved with a non-blocking follow-up: preserve a successful profile response when the optional parallel search fails, while retaining existing query-only failure behavior. The search threshold and limit change is accepted; no adapter code was changed as part of this review.
…ppers (#1799) Bundles 10 community fixes for the Python wrappers, each commit kept with its original author. Ships supermemory-openai-sdk 1.0.10, supermemory-cartesia 0.1.5 and supermemory-pipecat 0.1.5. - openai-sdk-python: shared OpenAI clients no longer leak memories across tenants (#1503), memory search fails open instead of crashing the chat (#1288), the requests fallback runs off the event loop (#1660), get_profile works for users with no memories (#1759). - cartesia: the wrapped agent runs exactly once per event (#1280, #1670), plain-string and chunk-only search results are kept (#1683, #1693). - pipecat: memory text stays literal when replacing the injected block (#1647), `InputParams` is exported from the package root (#1575). - #1503 was hand-ported onto main's newer middleware. #1751 is left out: the v5 agent-framework migration (#1786) appends to the conversation document by design. Co-Authored-By: abhinav7x94 <204053250+abhinav7x94@users.noreply.github.com> Co-Authored-By: shamAnimates <145093437+shamAnimates@users.noreply.github.com> Co-Authored-By: aloktripathi1 <142301030+aloktripathi1@users.noreply.github.com> Co-Authored-By: Cintu07 <178455858+Cintu07@users.noreply.github.com> Co-Authored-By: Atharvsinh-codez <201381760+Atharvsinh-codez@users.noreply.github.com> Co-Authored-By: abhay-codes07 <182421137+abhay-codes07@users.noreply.github.com> Co-Authored-By: nkootstra <545768+nkootstra@users.noreply.github.com> Co-Authored-By: Agnik47 <140933190+Agnik47@users.noreply.github.com>
Prepares
supermemory-agent-framework2.0.0 for the Supermemory Python SDK>=5.0.0,<6. Provider, middleware, and tools use namespace-first ingestion/search/profile calls; full-mode retrieval fetches profile and search separately. Writes retain POST append/diff behavior.Preserves connection arguments, exported names, conversation document IDs, tool JSON envelopes, string-based profile facts/buckets, top-level update timestamps, and scoped session state. Legacy optional search fields unavailable from v5 remain present as
null, rather than preserving their old values. Existing server-side data availability in v5 is not verified by the mocked tests; the README and integration docs retain that migration caveat.README, integration docs, docstrings, and interactive examples use
OpenAIChatClient(model="gpt-5"). Normal pip/uv setup requiressupermemory-agent-framework>=2.0.0,<3and includesagent-framework-openai, replacing temporary source-checkout instructions without accidentally installing the legacy adapter. Agent Framework CI SDK matrix lanes now use 5.0.0; other adapters and publishing/deployment workflows are unchanged.Validated all 72 tests with SDK 5.0.0 against framework cores 1.0.0rc3 and 1.21.0, including actual SDK HTTP decoding and framework agent/tool loops under mocked transport. Built the 2.0.0 wheel, verified its metadata, installed it into a clean environment, and reran all 72 tests against the installed wheel. Type/format/lint and dependency checks pass. All 21 README/docs Python examples parse, MDX compiles, and
gpt-5client/agent construction succeeds with network access blocked. Live verification now also exercised real Supermemory and OpenAI (gpt-5-2025-08-07) services using the built wheel. The model invokedadd_memory, all provider/middleware/tool writes mapped to the unchanged conversation document ID, and source content was persisted.Live processing caveat: the unchanged dynamic-write flow produced no learned memories/profile within 30 minutes in a fresh namespace, despite document status
done, so its fresh-session recall did not pass. Direct SDK controls withdreaming="instant"produced learned memories and populated profiles. In a separate control using an explicitly instant-seeded document, the unchanged adapter retrieved those facts into a fresh GPT-5 session, which recalled both synthetic preferences; another empty namespace returnedUNKNOWNand no facts. That passing control does not establish that the adapter’s default write-to-recall flow passed. All test documents/memories were deleted and empty search/profile results verified. No adapter defaults or billing behavior were changed.Release behavior: 2.0.0 was unused on PyPI when checked. Merging the
pyproject.tomlchange intomainautomatically triggers the existing Publish Agent Framework Python workflow; publication and the main-branch Mintlify deployment run independently. The PR docs preview deployed successfully, but 2.0.0 has not been published and production docs are unchanged by this branch. landing-2#65 has already merged and deployed; its adapter 2.0.0 installation requirement depends on this package release. This adapter PR has not been merged or published by this thread.