Names like constructor, toString and __proto__ resolved Object.prototype members as schemas and threw on parameters.required, failing the whole tool batch. The schema lookup now uses Object.hasOwn, so they return the usual unknown-function error.
- New claude-memory IDs use a cm_ prefix so they can never match a legacy ID, and a write refuses an ID held by another document.
- rename overwrites an existing destination (including a legacy ID) instead of leaving a stale copy.
- Non-404 lookup failures return an error instead of allowing a write; delete works while content is still processing.
- OpenAI executor rejects blank or missing required arguments, and documentDelete requires a documentId.
- MemoryCommand.path is required again for every command except rename.
Ports the #1676 collision fix to v5 with a reversible path ID and a legacy-ID fallback, so files written by 3.1.1 still open and get overwritten in place (also while their content is still processing). Keeps new_str required for str_replace and moves the PR tests onto the v5 mocks.
`getProfile`, `documentList` and `memoryForget` all declare `required: []`,
so a model may legitimately call them with no arguments at all. The OpenAI
API serialises that as `arguments: ""`, and `parseToolArguments` handed the
empty string straight to `JSON.parse`, so every no-argument call came back as
{"success":false,"error":"Invalid JSON arguments for getProfile"}
Those three tools were unreachable in their documented no-argument form.
The same gate also lets non-object JSON through. `"null"` parses cleanly and
then rejects in the destructuring parameter of every tool function --
`TypeError: Cannot destructure property 'containerTag' of 'object null'` --
which escapes `executeToolCall`, since it has no catch, and fails the whole
request. That is precisely the throw #1488 added this gate to contain. `"5"`
and `"\"text\""` are quieter but worse: they destructure to `undefined` and
call the API with no container tag at all.
Treat blank arguments as `{}`, and require the parsed value to be a non-null,
non-array object. Malformed JSON still returns the tool error #1488 added.
Adds eight regression tests. Six of them fail against the current
implementation -- two on the blank-argument path and four on the non-object
path, one carrying the raw TypeError. The two guard tests, malformed JSON and
an ordinary well-formed call, pass both before and after, so the behaviour
Three places where ClaudeMemoryTool diverges from the documented
memory_20250818 wire format:
- rename sends old_path/new_path, not path. handleCommand validated
command.path, so every rename coming from a real model died with
"Cannot read properties of undefined (reading 'startsWith')".
path is still accepted as the source for existing callers.
- insert_line means "insert after this line" (0 = top of file), but we
spliced at insertLine - 1 and rejected 0, so every insert landed one
line above where Claude asked and inserting at the top was impossible.
- str_replace with new_str omitted is a deletion per the spec; we
rejected it.
The new tests mock the supermemory client so they run without an API
key. Also fixed the rename example in the docs, which showed the same
path shape the code expected.
Anthropic's memory_20250818 spec defines insert as: insert_text is inserted
AFTER line insert_line, 0 inserts at the beginning of the file, and the valid
range is [0, n_lines]. The implementation treated insert_line as a 1-based
insert-BEFORE index with range [1, n_lines + 1].
Since the caller of this tool is Claude itself, which is trained on the spec
semantics, every model-driven insert landed one line earlier than intended,
insert_line: 0 (insert at top of file) was rejected as invalid, and
insert_line: n_lines (append) inserted before the last line instead of after
it.
Fix the validation range to [0, n_lines], splice at insert_line directly
(0-based insert-after), and update the error and success messages to match.
One existing tool-operations test encoded the old insert-before behavior; its
insert_line is adjusted so its expected output is unchanged under spec
semantics. Adds four regression tests covering top-of-file, middle,
append, and both out-of-range directions.
Every integration (AI SDK, OpenAI, Mastra, VoltAgent, Claude memory) now calls v5 through `supermemory@5`. The config takes one `namespace` instead of `containerTags` or `projectId`, and `withSupermemory` takes `namespace` and `id` instead of `containerTag` and `customId`. No aliases. With no config the tools still use `sm_project_default`.
Conversations are stored as one document per conversation keyed by `id` instead of `/v4/conversations`. `memoryForget` drops `reason`; forgetting by text previews with `forgetMatching` and then forgets exact matches by id. Claude memory marks its files with `metadata.source` instead of a second tag. Search keeps the 2.x defaults so results do not shift.
`packages/ai-sdk` is removed: npm already deprecates it in favor of `@supermemory/tools/ai-sdk`, so its CI steps and playground wiring go too. `apps/sdk-playground` moves to the new tools API and v5 routes, including its Python server.
Tested against production: add, search, profile, list and delete. 112 unit tests pass; type errors drop from 148 to 141, none new. The `supermemory` dependency pins the rc until 5.0.0 is published.
## Stack Context
This stack moves memory deduplication **out of the playground UI and into the SDKs themselves**, so every integration injects a single, deduplicated, self-replacing memory block. Three PRs:
1. **`sdk-dedup/tools-ts`** (this PR) — TypeScript SDK core + integrations
2. `sdk-dedup/python` — Python SDKs
3. `sdk-dedup/playground` — playground debug view reflects the SDK-owned block
## What?
Move profile deduplication into the SDK middleware for the TypeScript tools package.
- Facts are normalized (strip leading `[YYYY-MM-DD]`, trim, collapse whitespace, casefold) and deduplicated in **`static > dynamic > search`** priority within a single request.
- The result is injected as one **owned `<supermemory>` block** that *replaces* the previous block instead of accumulating a new one each turn.
- Dedup is **mode-aware**: in query mode, search results are not dropped against a profile that isn't being injected.
- Deduplication is **request-local** — no global/browser `Set`. Safe for multiple users, concurrent requests, and Cloudflare Worker isolates.
Covers AI SDK, OpenAI (Chat + Responses), Mastra, and VoltAgent. New `shared/memory-context.ts` owns the block-replacement logic.
## Why?
The earlier "conversation-scoped deduplication" was only a playground browser `Set` — a UI debug affordance that did not change what the SDK sent to the model, and would have been unsafe as server-side global state. Real cross-source dedup belongs in the SDK, applied fresh per stateless model request.
## Testing
- `bun run test` in `packages/tools`: 145 passed (the one failing suite, `claude-memory.test.ts`, is a pre-existing broken import unrelated to this change).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> **Medium Risk**
> Changes how system prompts and instructions are built across all TypeScript integrations; behavior is well-covered by unit tests but incorrect strip/replace logic could drop or duplicate context in production prompts.
>
> **Overview**
> Moves **cross-source memory deduplication** and **owned prompt injection** into `@supermemory/tools` so every integration sends one deduplicated memory block per request instead of growing context each turn.
>
> **Deduplication:** Facts are normalized via `normalizeMemoryFact` (strip `[YYYY-MM-DD]`, trim, collapse whitespace, lowercase) and deduplicated with **static → dynamic → search** priority. `deduplicateMemoriesForMode` keeps search hits in **query** mode when the profile is not injected.
>
> **Owned `<supermemory>` block:** New `shared/memory-context.ts` wraps memories in `<supermemory context="user-memories" readonly>`, strips stale blocks, and **replaces** prior SDK context while preserving caller system instructions. Applied in AI SDK (`injectMemoriesIntoParams`), OpenAI Chat/Responses middleware, Mastra input processor (`wrapMemoryContext`), and VoltAgent hooks.
>
> **Tests:** Unit coverage for block replacement (with-supermemory, OpenAI, VoltAgent), Mastra wrapper tag assertion, normalized dedup variants, and concurrent `containerTag` isolation.
>
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit 2fa2e0d85c. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
## Summary
- Refresh canonical tool descriptions in `tools-shared.ts`
- Align OpenAI and AI SDK tool bindings with 7-tool surface
- Export `TOOL_DESCRIPTIONS` / `PARAMETER_DESCRIPTIONS` from package index
Stacked on #1431
## Test plan
- [ ] `bun run test:unit` in `packages/tools`
Made with [Cursor](https://cursor.com)
Cherry-picks four contributor PRs for `@supermemory/tools` onto one branch, and bumps the package to 2.2.0.
- #1244 (@rajarshidattapy): `withSupermemory` accepts `options.apiKey` instead of only reading `SUPERMEMORY_API_KEY`, matching the Vercel, Mastra and Voltagent integrations. Unblocks secrets managers, edge runtimes and per-request keys.
- #1574 (@Agnik47): re-exports `PromptTemplate`, `MemoryPromptData` and `WithSupermemoryOptions` from `ai-sdk`. `./vercel` is not a published subpath, so the documented custom-template example did not compile.
- #1488 (@abhinav7x94): malformed tool-call JSON returns an error result instead of throwing out of the request.
- #1507 (@abhinav7x94): VoltAgent `onEnd` awaits the conversation save, which was fire-and-forget and could be dropped when a serverless runtime tore down.
Dropped the `middleware.test.ts` added by #1244. Note that editing `packages/tools/package.json` triggers the npm publish workflow on merge.
Co-Authored-By: rajarshidattapy <138959719+rajarshidattapy@users.noreply.github.com>
Co-Authored-By: Agnik47 <140933190+Agnik47@users.noreply.github.com>
Co-Authored-By: abhinav7x94 <204053250+abhinav7x94@users.noreply.github.com>
**`withSupermemory`** **(AI SDK)**
- **`skipMemoryOnError`** **defaults to** **`true`**. memory errors/timeouts log and the model runs on the **original** prompt unless you set `skipMemoryOnError: false`.
- **Pre-LLM** **`/v4/profile`** **is aborted after 5s** via `AbortSigna`
**Docs**
- `packages/tools/README.md`, **`apps/docs/integrations/ai-sdk.md`**
adds withSupermemory wrapper and input/output processors for
mastra agents:
- input processor fetches and injects memories into system prompt
before llm calls
- output processor saves conversations to supermemory after
responses
- supports profile, query, and full memory search modes
- includes custom prompt templates and requestcontext support
const agent = new Agent(withSupermemory(
{ id: "my-assistant", model: openai("gpt-4o"), instructions:
"..." },
"user-123",
{ mode: "full", addMemory: "always", threadId: "conv-456" }
))
includes docs as well
this pr also reworks how the tools package works into shared modules
## Add customizable prompt templates for memory injection
**Changes:**
- Add `promptTemplate` option to `withSupermemory()` for full control over injected memory format (XML, custom branding, etc.)
- New `MemoryPromptData` interface with `userMemories` and `generalSearchMemories` fields
- Exclude `system` messages from persistence to avoid storing injected prompts
- Add JSDoc comments to all public interfaces for better DevEx
**Usage:**
```typescript
const customPrompt = (data: MemoryPromptData) => `
<user_memories>
${data.userMemories}
${data.generalSearchMemories}
</user_memories>
`.trim()
const model = withSupermemory(openai("gpt-4"), "user-123", {
promptTemplate: customPrompt,
})
```
