A product comparison starts with specifications. A useful business decision also needs context: the team's budget, warranty expectations, repair experiences, and requirements learned from previous purchases.

REcall combines web research with persistent team memory to compare work laptops, office monitors, and meeting headsets. It shows recommendations with and without remembered context, using the same research evidence, so the influence of memory is visible.

People make the decision. REcall does not buy products or place orders.

Demo access: The hosted application requires a fresh, private workspace invitation from the owner. A public page loading successfully does not mean research services are available. See hosted access before presenting.

Actual hosted interface captured on September 30, 2026. This image shows the comparison form, not a completed research result.

These are unedited interface captures from the deployed application. They document the visible UI state, not an end-to-end service certification.

flowchart TD
    USER[Team member] --> UI[Next.js workspace UI]
    INVITE[Private signed invitation] --> SESSION[Session route]
    SESSION --> COOKIE[HttpOnly workspace cookie]
    COOKIE --> UI
    UI --> API[Authenticated agent route]
    API --> GUARD[Origin, session, size, schema and rate checks]
    GUARD --> AGENT[Research and memory orchestration]

    AGENT <--> MEMORY[Hindsight team memory banks]
    AGENT --> SEARCH{Research provider}
    SEARCH -->|Gemini or Groq| TAVILY[Tavily web research]
    SEARCH -->|OpenAI| WEB[Responses web search]
    TAVILY --> EVIDENCE[Shared web evidence]
    WEB --> EVIDENCE
    EVIDENCE --> MODEL[Selected model provider]
    MEMORY -->|Recalled preferences and decisions| MODEL
    MODEL --> VIEWS[With-memory and without-memory summaries]
    VIEWS --> VALIDATE[Schema, product and citation validation]
    VALIDATE -->|Accepted| RESULT[Comparison, sources and uncertainties]
    RESULT --> UI

    classDef client fill:#eaf2ff,stroke:#3b6fba,color:#162b42
    classDef server fill:#edf7f2,stroke:#43866a,color:#173e2e
    classDef external fill:#fff5e5,stroke:#b88736,color:#553d15
    class UI,USER,RESULT client
    class SESSION,COOKIE,API,GUARD,AGENT,VALIDATE,VIEWS,EVIDENCE server
    class MEMORY,TAVILY,WEB,MODEL external

One comparison request: recall relevant memory → retrieve web evidence → generate and validate the with-memory view → generate and validate the without-memory view → return both views with sources.

Both views share the request's evidence. Invalid generated output gets at most one regeneration attempt per view, using the same supplied evidence; repeated invalid output is rejected. Provider failures retain their own error meaning rather than being treated as validation failures.

flowchart LR
    A[Human confirms a requirement or decision] --> B[Retain in team memory]
    B --> C[Receive save confirmation]
    C --> D[Return later and recall context]
    D --> E[Apply context to new research]
    E --> A

Memory captures preferences and experiences; it is not evidence of current prices or specifications. A saved correction adds context rather than guaranteeing deletion of an earlier record.

REcall/
├── app/
│   ├── page.tsx                    # Application entry point
│   └── api/
│       ├── agent/route.ts          # Compare, remember, seed, history, verify
│       └── session/route.ts        # Workspace invitation/session handling
├── components/
│   ├── recall-workspace.tsx        # Comparison, memory and status screens
│   └── ui/                        # Reusable interface components
├── lib/recall/
│   ├── core.ts                    # Schemas and comparison validation
│   ├── generation.ts              # Bounded regeneration and safe diagnostics
│   ├── live.ts                    # Research, model and Hindsight integration
│   └── session.ts                 # Signed access and process-local limits
├── data/team-history.json         # Eight labelled synthetic example records
├── public/                        # Favicon and README evidence images
├── scripts/invite.mjs             # Private workspace invitation generator
├── tests/generation.test.cjs      # Nine mocked generation regression tests
├── .env.example                   # Server-side configuration template
├── ARCHITECTURE.md                # Data flow and reliability boundaries
└── DEPLOYMENT.md                  # Hosting, private access and verification

Use Node.js 22.12 or later and npm.

git clone https://github.com/surya-3699/REcall.git
cd REcall
npm ci

Copy .env.example to .env.local using your editor or file manager. Keep credentials server-side; never commit .env.local or put secrets in NEXT_PUBLIC_* variables.

For the Gemini path, configure Hindsight, Gemini, and Tavily. Choose a model your API project can access; the template's model name is a project default, not a guarantee of availability or quota.

Leave unused provider keys blank. API access, billing, and rate limits belong to the configured provider accounts; setting a key does not itself establish usable capacity.

npm run dev

Open localhost:3000. Loopback development can establish a local session automatically. Hosted access uses signed invitations instead.

1. Open Service status and select Check research services.
2. Choose a team, then enter two exact product variants, the market, and requirements.
3. Run a comparison. Inspect linked evidence, unknowns, and the with-memory/without-memory views.
4. Under Team memory, save a confirmed requirement or decision and wait for the save receipt.
5. Reload the page, keep the same workspace/team, and refresh memory.
6. Run a later comparison and inspect whether the recalled context is used appropriately.

Example history is synthetic, not a record of real purchasing approvals or current product facts. Loading it again uses stable document IDs. Memory refresh retrieves relevant context; it is not a complete comparison archive.

Deploy as a Node.js Next.js application over HTTPS. Put provider credentials and SESSION_SECRET in the host's private environment settings. Ensure the host can reach external services and supports the request duration; no particular free tier or hosting plan is guaranteed.

The owner generates an invitation from a private local .env.local containing the same signing secret as the deployment:

npm run invite -- https://r-ecall.vercel.app judges

Replace the URL if deploying elsewhere. judges is the workspace ID; use the same ID when people should share workspace memory. Invitations expire after 15 minutes and successful access creates a seven-day browser session. Use the intended deployment domain consistently.

Share invitations privately, never in this README, an issue, or a public recording. See the deployment guide for setup and verification.

npm test
npm run typecheck
npm run build

The regression suite covers valid output, unsupported citations, uncited claims/reviews, incorrect products, fabricated memory references, bounded retries, malformed JSON diagnostics, and provider-error propagation.

• Provider credentials stay on the server. Hosted routes require a signed workspace session and same-origin requests; session cookies are HttpOnly and use Secure on hosted deployments.
• Request schemas and a 16 KB JSON limit constrain accepted input. Citation-ID validation checks references, not whether every cited claim is factually correct.
• The rate limiter is process-local: it is not durable or shared across replicas. Add shared quota enforcement before relying on multi-instance limits.
• The agent has a 170-second overall budget; the route advertises a 240-second maximum. Hosting/provider limits may end requests earlier.
• Supported transient Gemini failures may use fallback models and an explicitly configured Groq backup. An evidence-only preview during an outage is not a completed AI recommendation. Authentication and quota errors may still fail the request.
• Web excerpts can be stale, incomplete, or refer to another regional variant. Review original sources before buying. The paired views illustrate context use; they are not a controlled quality benchmark.

Keep changes focused, avoid committing credentials, and run the verification commands before proposing a change. Include screenshots for interface changes and explain any effects on memory, evidence handling, or provider configuration.

• Hindsight source and documentation
• System architecture and deployment guide
• MIT license — copyright 2026 REcall Contributors