feat(tools): mass forget + bucket tools, and refresh the supermemory skill - #1453
Conversation
…skill @supermemory/tools had no tool for POST /v4/memories/forget-matching or bucketed profile reads, so agents could forget one memory at a time and read the whole profile or nothing. Adds both, mirrored across the AI SDK and OpenAI surfaces, over fetch — the generated SDK has no forget-matching method and profile()'s params can't express include/buckets. memoryForgetMatching defaults dryRun to true even though the API defaults to false: a model shouldn't be able to bulk-delete in one turn. The Claude skill had drifted further. profile() was called with instead of , the profile response shape was invented, search responses were treated as arrays, Python used dict access on Pydantic models, and filters used a form the API rejects. The reference documented 3 endpoints and never mentioned forgetting at all. Rewrote it against the route schemas in mono, with a table marking which endpoints have SDK methods and which need HTTP.
Deploying with
|
| Status | Name | Latest Commit | Preview URL | Updated (UTC) |
|---|---|---|---|---|
| ✅ Deployment successful! View logs |
supermemory-app | dfddc2a | Commit Preview URL | Aug 10 2026, 01:00 PM |
Deploying with
|
| Status | Name | Latest Commit | Updated (UTC) |
|---|---|---|---|
| ✅ Deployment successful! View logs |
supermemory-mcp | dfddc2a | Aug 10 2026, 12:59 PM |
How to use the Graphite Merge QueueAdd the label Main to this PR to add it to the merge queue. You must have a Graphite account in order to use the merge queue. Sign up using this link. An organization admin has enabled the Graphite Merge Queue in this repository. Please do not merge from GitHub as this will restart CI on PRs being processed by the merge queue. This stack of pull requests is managed by Graphite. Learn more about stacking. |
|
Claude finished @sohamd22's task in 2m 7s —— View job Review complete for PR #1453
Overview: Adds Issues found: None — this looks good to ship. The new tools are well-implemented with proper safety defaults ( Score: 10/10 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Overview: Adds memoryForgetMatching and getProfileBuckets tools to @supermemory/tools, and fixes significant documentation drift in the supermemory skill (incorrect param names, invented response shapes, invalid filter syntax).
Issues found: None — this looks good to ship.
Details:
The new tools are well-implemented:
memoryForgetMatchingTooldefaultsdryRun: truewhich is the right safety default for a bulk-delete operation- Input validation requires at least one of
queryormemoryIdsbefore making the request - Both AI SDK and OpenAI surfaces are covered with consistent implementations
- Good test coverage including the dry-run default behavior and validation
The skill documentation fixes address real problems:
query→qfor the profile endpoint- Invented
context.profile/context.memories→ actual{ profile: { static, dynamic }, searchResults }shape - Python dict access → Pydantic model attribute access
- Bare
{ metadata: {...} }filters → requiredAND/ORwrapper - Added coverage for endpoints that were never documented (forget-matching, buckets, conversations, versioning)
Score: 10/10
1 participant
@supermemory/toolshad no tool forPOST /v4/memories/forget-matchingor bucketed profile reads, so agents could forget one memory at a time and read the whole profile or nothing. Adds both —memoryForgetMatchingTool/createMemoryForgetMatchingFunctionandgetProfileBucketsTool/createGetProfileBucketsFunction— mirrored across the AI SDK and OpenAI surfaces, overfetch: the generated SDK has noforget-matchingmethod andprofile()'s params can't expressinclude/buckets.memoryForgetMatchingdefaultsdryRuntotrueeven though the API defaults tofalse: a model shouldn't be able to bulk-delete in one turn. The intended loop is preview → show the user the returnedmemories→ call again with thosememoryIdsanddryRun: false.The Claude skill had drifted further.
profile()was called withqueryinstead ofq, the profile response shape was invented (context.profile,context.memories), search responses were treated as flat arrays instead of{ results, total, timing }, Python used dict access on Pydantic models, and filters used a bare{ metadata: {...} }form the API rejects. The reference documented 3 endpoints and never mentioned forgetting, versioning, conversations, container tags, or buckets at all. Rewrote it against the actual route schemas, with a table marking which endpoints have SDK methods and which need a plain HTTP call, and added a shared helper pattern for the HTTP-only paths. Quickstart and use-case examples were updated to match the real response shapes throughout.