[Feature] Implementation of token buffer memory

docs/docs/modules/memory/how_to/buffer_token_memory.mdx

@@ -0,0 +1,20 @@
+# Conversation token buffer memory
+
+This notebook covers how to use `ConversationTokenBufferMemory`. This memory keeps a buffer of recent interactions in memory, and uses token length rather than number of interactions to determine when to flush interactions.
+
+Below is a basic demonstration of the usage of token buffer memory.
+
+import CodeBlock from "@theme/CodeBlock";
+import Example from "@examples/memory/token_buffer.ts";
+
+<CodeBlock language="typescript">{Example}</CodeBlock>
+
+We can also get the history as a list of messages, useful if you are using this with `MessagesPlaceholder` in a chat prompt template.
+
+```typescript
+const memory = new ConversationTokenBufferMemory({
+  llm: model,
+  maxTokenLimit: 10,
+  returnMessages: true,
+});
+```

examples/src/memory/token_buffer.ts

@@ -0,0 +1,18 @@
+import { OpenAI } from "langchain/llms/openai";
+import { ConversationTokenBufferMemory } from "langchain/memory";
+
+const model = new OpenAI({});
+const memory = new ConversationTokenBufferMemory({
+  llm: model,
+  maxTokenLimit: 10,
+});
+
+await memory.saveContext({ input: "hi" }, { output: "whats up" });
+await memory.saveContext({ input: "not much you" }, { output: "not much" });
+
+const result1 = await memory.loadMemoryVariables({});
+console.log(result1);
+
+/*
+  { history: 'Human: not much you\nAI: not much' }
+*/

langchain/src/memory/buffer_token_memory.ts

@@ -0,0 +1,115 @@
+import {
+  InputValues,
+  MemoryVariables,
+  getBufferString,
+  OutputValues,
+} from "./base.js";
+
+import { BaseChatMemory, BaseChatMemoryInput } from "./chat_memory.js";
+import { BaseLanguageModel } from "../base_language/index.js";
+
+/**
+ * Interface for the input parameters of the `BufferTokenMemory` class.
+ */
+
+export interface ConversationTokenBufferMemoryInput
+  extends BaseChatMemoryInput {
+  /* Prefix for human messages in the buffer. */
+  humanPrefix?: string;
+
+  /* Prefix for AI messages in the buffer. */
+  aiPrefix?: string;
+
+  /* The LLM for this instance. */
+  llm: BaseLanguageModel;
+
+  /* Memory key for buffer instance. */
+  memoryKey?: string;
+
+  /* Maximmum number of tokens allowed in the buffer. */
+  maxTokenLimit?: number;
+}
+
+/**
+ * Class that represents a conversation chat memory with a token buffer.
+ * It extends the `BaseChatMemory` class and implements the
+ * `ConversationTokenBufferMemoryInput` interface.
+ */
+
+export class ConversationTokenBufferMemory
+  extends BaseChatMemory
+  implements ConversationTokenBufferMemoryInput
+{
+  humanPrefix = "Human";
+
+  aiPrefix = "AI";
+
+  memoryKey = "history";
+
+  maxTokenLimit = 2000; // Default max token limit of 2000 which can be overridden
+
+  llm: BaseLanguageModel;
+
+  constructor(fields: ConversationTokenBufferMemoryInput) {
+    super(fields);
+    this.llm = fields.llm;
+    this.humanPrefix = fields?.humanPrefix ?? this.humanPrefix;
+    this.aiPrefix = fields?.aiPrefix ?? this.aiPrefix;
+    this.memoryKey = fields?.memoryKey ?? this.memoryKey;
+    this.maxTokenLimit = fields?.maxTokenLimit ?? this.maxTokenLimit;
+  }
+
+  get memoryKeys() {
+    return [this.memoryKey];
+  }
+
+  /**
+   * Loads the memory variables. It takes an `InputValues` object as a
+   * parameter and returns a `Promise` that resolves with a
+   * `MemoryVariables` object.
+   * @param _values `InputValues` object.
+   * @returns A `Promise` that resolves with a `MemoryVariables` object.
+   */
+  async loadMemoryVariables(_values: InputValues): Promise<MemoryVariables> {
+    const messages = await this.chatHistory.getMessages();
+    if (this.returnMessages) {
+      const result = {
+        [this.memoryKey]: messages,
+      };
+      return result;
+    }
+    const result = {
+      [this.memoryKey]: getBufferString(
+        messages,
+        this.humanPrefix,
+        this.aiPrefix
+      ),
+    };
+    return result;
+  }
+
+  /**
+   * Saves the context from this conversation to buffer. If the amount
+   * of tokens required to save the buffer exceeds MAX_TOKEN_LIMIT,
+   * prune it.
+   */
+  async saveContext(inputValues: InputValues, outputValues: OutputValues) {
+    await super.saveContext(inputValues, outputValues);
+
+    // Prune buffer if it exceeds the max token limit set for this instance.
+    const buffer = await this.chatHistory.getMessages();
+    let currBufferLength = await this.llm.getNumTokens(
+      getBufferString(buffer, this.humanPrefix, this.aiPrefix)
+    );
+
+    if (currBufferLength > this.maxTokenLimit) {
+      const prunedMemory = [];
+      while (currBufferLength > this.maxTokenLimit) {
+        prunedMemory.push(buffer.shift());
+        currBufferLength = await this.llm.getNumTokens(
+          getBufferString(buffer, this.humanPrefix, this.aiPrefix)
+        );
+      }
+    }
+  }
+}

langchain/src/memory/index.ts

@@ -35,3 +35,7 @@ export {
   ConversationSummaryBufferMemory,
   type ConversationSummaryBufferMemoryInput,
 } from "./summary_buffer.js";
+export {
+  ConversationTokenBufferMemory,
+  type ConversationTokenBufferMemoryInput,
+} from "./buffer_token_memory.js";

langchain/src/memory/tests/buffer_token_memory.int.test.ts

@@ -0,0 +1,53 @@
+import { test, expect } from "@jest/globals";
+import { OpenAI } from "../../llms/openai.js";
+import { ConversationTokenBufferMemory } from "../buffer_token_memory.js";
+import { ChatMessageHistory } from "../../stores/message/in_memory.js";
+import { HumanMessage, AIMessage } from "../../schema/index.js";
+
+test("Test buffer token memory with LLM", async () => {
+  const memory = new ConversationTokenBufferMemory({
+    llm: new OpenAI(),
+    maxTokenLimit: 10,
+  });
+  const result1 = await memory.loadMemoryVariables({});
+  expect(result1).toStrictEqual({ history: "" });
+
+  await memory.saveContext({ input: "foo" }, { output: "bar" });
+  const expectedString = "Human: foo\nAI: bar";
+  const result2 = await memory.loadMemoryVariables({});
+  expect(result2).toStrictEqual({ history: expectedString });
+
+  await memory.saveContext({ foo: "foo" }, { bar: "bar" });
+  await memory.saveContext({ foo: "bar" }, { bar: "foo" });
+  const expectedString3 = "Human: bar\nAI: foo";
+  const result3 = await memory.loadMemoryVariables({});
+  expect(result3).toStrictEqual({ history: expectedString3 });
+});
+
+test("Test buffer token memory return messages", async () => {
+  const memory = new ConversationTokenBufferMemory({
+    llm: new OpenAI(),
+    returnMessages: true,
+  });
+  const result1 = await memory.loadMemoryVariables({});
+  expect(result1).toStrictEqual({ history: [] });
+
+  await memory.saveContext({ foo: "bar" }, { bar: "foo" });
+  const expectedResult = [new HumanMessage("bar"), new AIMessage("foo")];
+  const result2 = await memory.loadMemoryVariables({});
+  expect(result2).toStrictEqual({ history: expectedResult });
+});
+
+test("Test buffer token memory with pre-loaded history", async () => {
+  const pastMessages = [
+    new HumanMessage("My name's Jonas"),
+    new AIMessage("Nice to meet you, Jonas!"),
+  ];
+  const memory = new ConversationTokenBufferMemory({
+    llm: new OpenAI(),
+    returnMessages: true,
+    chatHistory: new ChatMessageHistory(pastMessages),
+  });
+  const result = await memory.loadMemoryVariables({});
+  expect(result).toStrictEqual({ history: pastMessages });
+});