LangChain | Brain Docs

What you wire

LangChain's memory abstraction has two hooks:

loadMemoryVariables — called before the LLM runs. Back it with recall to fetch the memories that answer the current input, and format them into the history variable your prompt template expects.
saveContext — called after the LLM responds. Back it with encode to persist the turn so future calls can recall it.

Because recall returns a membership verdict (Single / Many / None) and not a token-budgeted block, the adapter joins the recalled memory texts itself. On None there is nothing to inject and history is empty.

TypeScript

import { BrainHttpClient } from '@brain-db/sdk';
import { BaseChatMemory } from 'langchain/memory';

class BrainMemory extends BaseChatMemory {
  private brain = new BrainHttpClient({ apiKey: process.env.BRAIN_API_KEY! });

  get memoryKeys() {
    return ['history'];
  }

  async loadMemoryVariables(input: { input: string }) {
    const answer = await this.brain.recall({ query: input.input, max_results: 6 });
    const history =
      answer.answer_kind === 'none' ? '' : answer.memories.map((m) => `- ${m.text}`).join('\n');
    return { history };
  }

  async saveContext(input: { input: string }, output: { output: string }) {
    await this.brain.encode({ text: input.input });
    await this.brain.encode({ text: `assistant: ${output.output}` });
  }
}

Use it like any other BaseChatMemory:

const chain = new ConversationChain({ llm: model, memory: new BrainMemory() });

Python

from brain_db_sdk import BrainHttpClient
from langchain.memory.chat_memory import BaseChatMemory


class BrainMemory(BaseChatMemory):
    brain = BrainHttpClient(os.environ["BRAIN_API_KEY"])

    @property
    def memory_variables(self):
        return ["history"]

    def load_memory_variables(self, inputs):
        answer = self.brain.recall(inputs["input"], max_results=6)
        history = (
            ""
            if answer.answer_kind == "none"
            else "\n".join(f"- {m.text}" for m in answer.memories)
        )
        return {"history": history}

    def save_context(self, inputs, outputs):
        self.brain.encode(inputs["input"])
        self.brain.encode(f"assistant: {outputs['output']}")

Using it as a retriever instead

If you want Brain to feed a RAG chain rather than act as chat memory, wrap recall in a retriever: call brain.recall({ query }) in _getRelevantDocuments (TS) / _get_relevant_documents (Python) and map each MemoryHit to a LangChain Document (pageContent = hit.text). The same membership shape applies — an empty document list means Brain returned None.

What you wire

TypeScript

Python

Using it as a retriever instead

On this page