fix(tools): stop serving stale memories in the Mastra input processor

SupermemoryInputProcessor kept its memory cache on the processor instance,
which lives as long as the agent, and the cache has no expiry. Any hit
returned memories from an earlier generation. In profile mode, the default,
the query text is empty, so every message in a thread shared one cache key:
the profile was fetched once and then frozen.

Mastra runs processInput once per generation, so unlike the Vercel and
VoltAgent middleware there is no tool-call loop for this cache to save calls
on. Remove it. The replaced test only covered identical messages, which is
why it never caught this.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Abhishek Sharma 2026-11-02 08:28:00 +05:30
parent 8b47bd17c2
commit 7d91fb4d55
2 changed files with 59 additions and 33 deletions

View file

@ -16,7 +16,6 @@ import {
createLogger,
normalizeBaseUrl,
validateApiKey,
MemoryCache,
buildMemoriesText,
extractQueryText,
wrapMemoryContext,
@ -51,7 +50,6 @@ interface ProcessorContext {
addMemory: "always" | "never"
logger: Logger
promptTemplate?: PromptTemplate
memoryCache: MemoryCache<string>
}
/**
@ -73,7 +71,6 @@ function createProcessorContext(
addMemory: options.addMemory ?? "always",
logger,
promptTemplate: options.promptTemplate,
memoryCache: new MemoryCache<string>(),
}
}
@ -140,8 +137,8 @@ export class SupermemoryInputProcessor implements Processor {
const { messages, messageList, requestContext } = args
// Mastra owns tagged system messages by tag. Clear the previous value on
// every invocation so empty, skipped, cached, fresh, and error paths cannot
// leave stale Supermemory context behind.
// every invocation so empty, skipped, fresh, and error paths cannot leave
// stale Supermemory context behind.
messageList.clearSystemMessages("supermemory")
try {
@ -159,19 +156,6 @@ export class SupermemoryInputProcessor implements Processor {
}
const effectiveThreadId = getEffectiveId(this.ctx, requestContext)
const turnKey = MemoryCache.makeTurnKey(
this.ctx.namespace,
effectiveThreadId,
this.ctx.mode,
queryText || "",
)
const cachedMemories = this.ctx.memoryCache.get(turnKey)
if (cachedMemories) {
this.ctx.logger.debug("Using cached memories", { turnKey })
messageList.addSystem(wrapMemoryContext(cachedMemories), "supermemory")
return messageList
}
this.ctx.logger.info("Starting memory search", {
namespace: this.ctx.namespace,
@ -190,7 +174,6 @@ export class SupermemoryInputProcessor implements Processor {
})
if (memories) {
this.ctx.memoryCache.set(turnKey, memories)
messageList.addSystem(wrapMemoryContext(memories), "supermemory")
this.ctx.logger.debug("Injected memories into system prompt", {
length: memories.length,

View file

@ -11,6 +11,7 @@ import {
profileBody,
requestJson,
requestPath,
searchBody,
} from "../v5-fetch"
import {
RequestContext,
@ -194,10 +195,14 @@ describe("SupermemoryInputProcessor", () => {
expect(systemCall?.args[1]).toBe("supermemory")
})
it("should use cached memories on second call with same message", async () => {
fetchMock.mockImplementation(async () =>
jsonResponse(createMockProfileResponse(["Cached memory"])),
)
it("should fetch fresh profile memories for each new message", async () => {
let callCount = 0
fetchMock.mockImplementation(async () => {
callCount++
return jsonResponse(
createMockProfileResponse([`Fact from call ${callCount}`]),
)
})
const processor = new SupermemoryInputProcessor({
namespace: TEST_CONFIG.namespace,
@ -206,29 +211,67 @@ describe("SupermemoryInputProcessor", () => {
mode: "profile",
})
const firstList = createMockMessageList()
await processor.processInput({
messages: [createMessage("user", "Hello")],
systemMessages: [],
messageList: firstList,
abort: vi.fn() as never,
retryCount: 0,
})
const secondList = createMockMessageList()
await processor.processInput({
messages: [createMessage("user", "What do you know about me?")],
systemMessages: [],
messageList: secondList,
abort: vi.fn() as never,
retryCount: 0,
})
expect(fetchMock).toHaveBeenCalledTimes(2)
const injected = secondList.calls.find((c) => c.method === "addSystem")
expect(injected?.args[0]).toContain("Fact from call 2")
expect(injected?.args[0]).not.toContain("Fact from call 1")
})
it("should fetch fresh search memories when the same message is repeated", async () => {
let callCount = 0
fetchMock.mockImplementation(async () => {
callCount++
return jsonResponse(searchBody([`Memory from call ${callCount}`]))
})
const processor = new SupermemoryInputProcessor({
namespace: TEST_CONFIG.namespace,
id: TEST_CONFIG.id,
apiKey: TEST_CONFIG.apiKey,
mode: "query",
})
const messages: MastraDBMessage[] = [createMessage("user", "Hello")]
const args1: ProcessInputArgs = {
await processor.processInput({
messages,
systemMessages: [],
messageList: createMockMessageList(),
abort: vi.fn() as never,
retryCount: 0,
}
})
await processor.processInput(args1)
expect(fetchMock).toHaveBeenCalledTimes(1)
const args2: ProcessInputArgs = {
const secondList = createMockMessageList()
await processor.processInput({
messages,
systemMessages: [],
messageList: createMockMessageList(),
messageList: secondList,
abort: vi.fn() as never,
retryCount: 0,
}
})
await processor.processInput(args2)
expect(fetchMock).toHaveBeenCalledTimes(1)
expect(fetchMock).toHaveBeenCalledTimes(2)
const injected = secondList.calls.find((c) => c.method === "addSystem")
expect(injected?.args[0]).toContain("Memory from call 2")
expect(injected?.args[0]).not.toContain("Memory from call 1")
})
it("should refetch memories for different user message", async () => {