hindsight-skills

AI coding agent skills for Hindsight — long-term memory for AI agents.

Works with Claude Code, Codex, Gemini CLI, Cursor, Kiro, and Factory Droid.

Most agents forget everything between sessions. Hindsight fixes that. It extracts facts, builds entity graphs, and synthesizes mental models from conversations — so your agent remembers users, learns patterns, and gets better over time.

These skills give your coding agent deep expertise in Hindsight’s architecture, so you can go from “I want my agent to remember things” to working code in a single session.

See it work

You:    /hindsight-architect

Claude: [reads README, package.json, scans directory structure, checks for LLM usage]

        I've looked at your project. It's a Next.js + Python customer support
        platform with OpenAI calls in the ticket handler and no existing
        Hindsight config.

        How do you want to approach adding memory?
        > A) Find opportunities for me
          B) I already know what I want
          C) Chat about it

You:    A

Claude: [examines handlers, routes, LLM calls, data flows]

        Here's how memory would work in this application:

        Memory flow: Retain full ticket conversations per-user, recall past
        context before each agent response, and synthesize patterns over time.

        Integration points:
        - api/chat.py:42 — add retain after each conversation turn
        - api/chat.py:18 — add recall before the LLM call
        - lib/agent.ts:31 — inject mental model content into system prompt

        What this enables: The agent remembers every user's history, learns
        each customer's product setup, and detects recurring issues across
        your entire base.

        Is this the direction you want, or would you adjust the scope?

You:    This is right. Let's do all of it.

Claude: Are there any of these considerations for your solution?
        > [ ] Enterprise security
          [ ] Data privacy / PII
          [ ] Regulatory compliance
          [x] None of these

Claude: [presents the three architecture decisions with reasoning]
        1. What to retain — full conversations, upsert on session_id
        2. Tag schema — userId:{id}, customerId:{id}
        3. Mental models — per-user preferences, per-customer usage, cross-customer issues

        Does this design work?

You:    Yes.

Claude: [asks about deployment: Cloud, Local, or Self-hosted]
        [asks about recall optimization: Balanced, Completeness, Cost, or Custom]

Claude: # Memory Implementation Plan: support-agent
        ## Bank Configuration — single bank, per-user tag isolation
        ## Tag Schema — userId:{id}, customerId:{id}
        ## Retain Strategy — full conversation upsert with document_id
        ## Recall Strategy — 10K token budget, any_strict tag matching
        ## Mental Models — 3 models with auto-refresh