Skip to main content
Hindsight is State-of-the-Art on Memory for AI Agents | Read the paper β†’
πŸ€–
Using a coding agent? Run this to install the Hindsight docs skill:
npx skills add https://github.com/vectorize-io/hindsight --skill hindsight-docs

Dify

Persistent memory for Dify workflows via Hindsight. The Hindsight Dify plugin adds three tools β€” Retain, Recall, Reflect β€” that work alongside any other Dify node in workflows, chatflows, and agent apps.

Why this matters​

Dify is one of the most popular open-source LLM app platforms β€” visual workflow builder, prompt management, and a growing tool/plugin ecosystem. Until now, Dify workflows have been stateless across runs. With Hindsight tools you can:

  • Retain every closed support ticket, sales-call transcript, or form submission into a memory bank
  • Recall relevant context before an LLM step so the model sees prior history
  • Reflect to ask synthesizing questions ("What do we know about this customer?") inside a workflow

Installation​

In your Dify dashboard go to Plugins β†’ Install Plugin, then choose one of:

  • Marketplace β€” search for Hindsight (once published)
  • GitHub β€” install from vectorize-io/hindsight (path hindsight-integrations/dify)
  • Local β€” upload the .difypkg archive

After install, the Hindsight plugin appears under Tools in the workflow editor.

Setup​

Recommended: Hindsight Cloud

Sign up free and grab an API key β€” no self-hosting required.

  1. Sign up at Hindsight Cloud (free tier) or self-host
  2. Get an API key from the Hindsight dashboard
  3. In Dify, open the Hindsight plugin and add credentials:
    • API URL β€” defaults to https://api.hindsight.vectorize.io (Cloud); change for self-hosted
    • API Key β€” your hsk_... key (optional for self-hosted unauthenticated instances)

Tools​

Retain​

Store content in a bank. Hindsight extracts facts asynchronously after the call returns.

FieldDescription
Bank IDMemory bank to store in (auto-created on first use)
ContentFree-text content to retain
TagsOptional comma-separated tags

Recall​

Search a bank for memories relevant to a query. Returns a results array.

FieldDescription
Bank IDMemory bank to search
QueryNatural-language query
Budgetlow / mid / high
Max TokensCap on returned tokens
TagsOptional tag filter (comma-separated)

Reflect​

Get an LLM-synthesized answer over the bank. Returns text.

FieldDescription
Bank IDMemory bank
QueryQuestion to answer
Budgetlow / mid / high

Example workflows​

Customer-support assistant β€” every closed Zendesk ticket retains the resolution. Every new ticket starts with a Recall against the bank to surface similar past issues, then passes that context to OpenAI/Anthropic to draft the first reply.

Sales-call coach β€” Gong webhook β†’ Hindsight Retain (call summary). Before each next prep call, Recall on the prospect's name to pull every prior touchpoint, then format into the daily prep doc.

Knowledge-base agent β€” every uploaded document is chunked and retained. The Dify chatflow uses Recall instead of vector-DB-only retrieval, getting fact-extracted, deduplicated, time-aware results.

Source​

AltStyle γ«γ‚ˆγ£γ¦ε€‰ζ›γ•γ‚ŒγŸγƒšγƒΌγ‚Έ (->γ‚ͺγƒͺγ‚ΈγƒŠγƒ«) /