@pensyve/langchain

Persistent AI memory for LangChain.js / LangGraph.js agents via Pensyve. Two complementary features:

1. Working-memory substrate — A system-prompt document (SUBSTRATE_PROMPT.md) that gives your LangGraph.js agent the reasoning discipline to recall before answering and capture lessons as they land.
2. Memory store backend — PensyveStore: a drop-in BaseStore-compatible backend backed by Pensyve's 8-signal fusion retrieval engine.

What It Does

The working-memory substrate is a reasoning layer — not a library — that you load into your agent's system prompt. Once loaded, the agent will:

• Recall before substantive answers using pensyve_recall, scoped by entity.
• Capture lessons in-flight using pensyve_observe when a root cause is confirmed, a decision lands, or an approach is abandoned.
• Manage episode lifecycle lazily: open an episode on the first observe, reuse it throughout the conversation.
• Surface memory lightly — one line per recall or capture, never narrating empty recalls.
• Wrap up sessions by presenting memory candidates for user confirmation before storage.

Install

bun add @langchain/anthropic @langchain/langgraph @langchain/mcp-adapters

# Memory store backend (optional — separate from the substrate)
bun add @pensyve/langchain

Set your API key:

export PENSYVE_API_KEY="psy_your_key_here"
export ANTHROPIC_API_KEY="sk-ant-..."

Create an API key at pensyve.com/settings/api-keys.

Quick Start

cd integrations/langchain-ts
bun run examples/pensyve-agent.ts

The example connects a LangGraph.js ReAct agent to the Pensyve MCP server and loads SUBSTRATE_PROMPT.md as the system prompt.

System Prompt

SUBSTRATE_PROMPT.md consolidates all eight substrate rules into a single document. Load it into your agent:

import { readFileSync } from "node:fs";
import { join } from "node:path";
import { createReactAgent } from "@langchain/langgraph/prebuilt";

const substrate = readFileSync(join(__dirname, "SUBSTRATE_PROMPT.md"), "utf-8");
const agent = createReactAgent({ llm, tools, prompt: substrate });

MCP Connection

import { MultiServerMCPClient } from "@langchain/mcp-adapters";

const client = new MultiServerMCPClient({
  pensyve: {
    transport: "streamable_http",
    url: "https://mcp.pensyve.com/mcp",
    headers: { Authorization: `Bearer ${process.env.PENSYVE_API_KEY}` },
  },
});
const tools = await client.getTools();
// ... use agent, then:
await client.close();

Memory Behavior Model

| Trigger | Action | MCP call | |---|---|---| | Before substantive answer | Recall by entity | pensyve_recall(query, entity, types, limit=5) | | Root cause confirmed | Capture episodic | pensyve_observe(episode_id, content, source_entity="langchain-ts", about_entity) | | Decision accepted | Capture semantic | pensyve_remember(entity, fact, confidence=0.9) | | Reusable workflow found | Capture procedural | pensyve_observe(... content="[procedural] ...") | | Session ending | Present candidates | User confirms before storage |

Memory Types

• Semantic — durable facts that remain true across sessions.
• Episodic — what happened in this thread (outcomes, root causes, abandoned approaches).
• Procedural — reusable workflows stored via pensyve_observe with a [procedural] prefix.

Opt-Out

To disable the substrate, remove SUBSTRATE_PROMPT.md from the agent's prompt argument. The PensyveStore backend is unaffected.

Links

• Pensyve — managed memory service
• API Keys
• LangChain.js docs
• LangGraph.js docs

License

Apache 2.0 — see LICENSE.